UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Automatic measurement of human subcutaneous fat with ultrasound Ng, Jessie Ying Chi 2006

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2006_fall_ng_jessie.pdf [ 1.99MB ]
Metadata
JSON: 831-1.0065566.json
JSON-LD: 831-1.0065566-ld.json
RDF/XML (Pretty): 831-1.0065566-rdf.xml
RDF/JSON: 831-1.0065566-rdf.json
Turtle: 831-1.0065566-turtle.txt
N-Triples: 831-1.0065566-rdf-ntriples.txt
Original Record: 831-1.0065566-source.json
Full Text
831-1.0065566-fulltext.txt
Citation
831-1.0065566.ris

Full Text

Automatic Measurement of Human Subcutaneous Fatwith UltrasoundbyJessie Ying Chi NgB.A.Sc., The University of British Columbia, 2002A THESIS SUBMITTED IN PARTIAL FULFILMENT OFTHE REQUIREMENTS FOR THE DEGREE OFMaster of Applied ScienceinThe Faculty of Graduate Studies(Electrical and Computer Engineering)The University of British ColumbiaSeptember 2006c© Jessie Ying Chi Ng 2006iiAbstractMeasuring human subcutaneous fat is useful for assessing health risks due to obesity and formonitoring athletes’ health status, body shapes and weight for various sports competitions such asgymnastics and wrestling. Our aim is to investigate the use of ultrasound imaging in automaticallymeasuring human subcutaneous fat thickness.We proposed to use the spectrum properties extracted from the raw radio frequency (RF) signalsof ultrasound for the purpose of fat boundary detection. Our fat detection framework consists offour main steps. The first step is capturing RF data from 11 beam steering angles and at fourfocal positions. Secondly, two spectrum properties (spectrum variance and integrated backscattercoefficient) are calculated from the local spectrum of RF data using the short time Fourier transformand moment analysis. The values of the spectrum properties are encoded as gray-scale parametricimages. Thirdly, spatial compounding is used to reduce speckle noise in the parametric imagesand improve the visualization of the subcutaneous fat layer. Finally, we apply Rosin’s thresholdingand Random Sample Consensus boundary detection on the parametric images to extract the fatboundary.The detection framework was tested on 36 samples obtained at the suprailiac, thigh and tricepsof nine human participants in vivo. When compared to manual boundary detection on ultrasoundimages, the best result was obtained from segmenting the spatial compounded spectrum variancevalues averaged over multiple focuses. A reasonable result could also be obtained by using a singlefocus. Further, our automatic detection results were compared with the results using skinfoldcaliper measurements. We found that the correlation is high between our automatic detectionand skinfold caliper measurement, and is similar to the previous studies which are not automatic.Our work has shown that the spatial compounded spectrum properties of RF data can be usedto segment the subcutaneous fat layer. Based on our results, it is feasible to detect fat at thesuprailiac, thigh and triceps sites using the spectrum variance. The values of spectrum variancechange more rapidly in the fat tissue than the non-fat tissue.iiiContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiContents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xGlossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviNotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxii1 Background and Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Current Techniques of Human Body Fat Measurement . . . . . . . . . . . . . . . . . 21.1.1 Body Density Weighing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.1.1 Underwater Weighing . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.1.2 Air-displacement Plethysmography . . . . . . . . . . . . . . . . . . . 31.1.2 Bioelectrical Impedance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 31.1.3 Skinfold Caliper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.1.4 Imaging Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.1.4.1 Dual Energy X-ray Absorptiometry . . . . . . . . . . . . . . . . . . 51.1.4.2 Computed Tomography . . . . . . . . . . . . . . . . . . . . . . . . . 51.1.4.3 Magnetic Resonance Imaging . . . . . . . . . . . . . . . . . . . . . . 51.1.4.4 Ultrasound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8Contents iv1.3 Pulse-Echo Ultrasound Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.3.1 Apparatus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.3.2 Interaction of Ultrasound with Soft Tissues . . . . . . . . . . . . . . . . . . . 151.3.3 Backscattered Radiofrequency Spectra . . . . . . . . . . . . . . . . . . . . . . 171.3.4 Current Development of Ultrasound Segmentation . . . . . . . . . . . . . . . 191.4 Properties of Human Subcutaneous Fat . . . . . . . . . . . . . . . . . . . . . . . . . 201.4.1 Biological characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201.4.2 Ultrasound characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201.4.3 Difficulties in Segmentation of Fat in Ultrasound Images . . . . . . . . . . . . 221.5 Thesis Objectives and Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 Method in Developing Image Processing and Boundary Detection . . . . . . . . 252.1 Processing of Radiofrequency Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.1.1 Calculation of Spectrum Properties . . . . . . . . . . . . . . . . . . . . . . . . 262.1.2 Noise Reduction using Spatial Compounding . . . . . . . . . . . . . . . . . . 282.1.2.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.1.2.2 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322.1.2.2.1 Phantom Experiments . . . . . . . . . . . . . . . . . . . . . 322.1.2.2.1.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . 322.1.2.2.1.2 Results and Discussions . . . . . . . . . . . . . . . . 332.1.2.2.2 Human Experiments . . . . . . . . . . . . . . . . . . . . . . 352.1.2.2.2.1 Method . . . . . . . . . . . . . . . . . . . . . . . . . 352.1.2.2.2.2 Results and Discussions . . . . . . . . . . . . . . . . 362.2 Thresholding on Spectrum Properties . . . . . . . . . . . . . . . . . . . . . . . . . . 402.2.1 Unimodal Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412.2.2 Thresholding results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422.3 Fat Boundary Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472.3.1 Extraction of Boundary Candidates . . . . . . . . . . . . . . . . . . . . . . . 472.3.2 Fitting Boundary Candidates using Random Sample Consensus (RANSAC) . 492.3.2.1 Total Least Squares Fitting . . . . . . . . . . . . . . . . . . . . . . . 50Contents v2.3.2.2 Theory and Implementation . . . . . . . . . . . . . . . . . . . . . . 502.3.3 Calculation of Spectral Content using Multiple Focuses . . . . . . . . . . . . 552.3.3.1 Stitching Focused Spectrum Properties (MF1) . . . . . . . . . . . . 552.3.3.2 Averaging Spectrum Properties from Multiple Focuses (MF2) . . . 562.3.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573 Experimental Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593.1 Overview of the Human Subcutaneous Fat Detection Framework . . . . . . . . . . . 593.1.1 Data capture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613.1.2 Calculation of Spectrum Properties . . . . . . . . . . . . . . . . . . . . . . . . 623.1.3 Pre-processing of Spectrum Properties Map . . . . . . . . . . . . . . . . . . . 633.1.4 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633.2 Procedures in User Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653.2.1 Measurement of Skinfold Fat Thickness . . . . . . . . . . . . . . . . . . . . . 653.2.2 Collection of Ultrasound Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 673.2.3 Reference Fat Boundary from Manual Segmentation . . . . . . . . . . . . . . 683.3 Evaluation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683.3.1 Average Thickness Error Metrics . . . . . . . . . . . . . . . . . . . . . . . . . 693.3.2 Root Mean Square Error Metric . . . . . . . . . . . . . . . . . . . . . . . . . 703.3.3 Difference against Mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704 Evaluation of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724.1 Qualitative Results: Segmentation Using Spectrum Variance σ2s . . . . . . . . . . . . 724.2 Evaluation of Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 754.2.1 Results: Spectrum Variance σ2s vs Integrated Backscattering Coefficient IBS 764.2.1.1 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 764.2.1.2 dERR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794.2.1.3 dRMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804.2.2 Results: Multiple-focuses vs Single Focuses . . . . . . . . . . . . . . . . . . . 81Contents vi4.2.2.1 dERR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814.2.2.2 dRMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 834.2.3 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 854.3 Comparison of Auto-detected Fat Thickness with Skinfold Caliper Measurements . . 934.3.1 Evaluation on Skinfold Caliper Technique . . . . . . . . . . . . . . . . . . . . 934.3.2 Result of Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 944.3.3 Result of Difference Against Mean . . . . . . . . . . . . . . . . . . . . . . . . 984.3.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 994.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1005 Conclusions and Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1025.1 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1025.2 Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117A Calculation of the Spectrum Central Frequency and Variance . . . . . . . . . . . 117B Normalization of a Spectrum Property Image . . . . . . . . . . . . . . . . . . . . . 119C Solution to Rosin’s Thresholding Method . . . . . . . . . . . . . . . . . . . . . . . . 120D UBC Research Ethics Board Certificate of Approval . . . . . . . . . . . . . . . . . 122viiList of Tables4.1 Average dERR of 36 samples in 9 subjects (4 samples per participant) at the suprail-iac, triceps and thigh sites respectively. The paired t-test is used to compare themean difference between σ2s and IBS. t(df=35): t-value of the paired t-test witha degree of freedom (df) of 35. CI : 95% confidence interval (CI) of the statisticalmean difference between σ2s and IBS. If the CI does not include 0, there is a sig-nificant difference between groups. p-value: if a p-value < 0.05, it indicates there issignificant difference between the the dERR of σ2s and IBS. . . . . . . . . . . . . . . 794.2 Average dRMS of 36 samples in 9 subjects (4 samples per participant) at the suprail-iac, triceps and thigh sites respectively. The paired t-test is used to compare themean difference between σ2s and IBS. t(df=35): t-value of the paired t-test witha degree of freedom (df) of 35. CI : 95% confidence interval (CI) of the statisticalmean difference between σ2s and IBS. If the CI does not include 0, there is a sig-nificant difference between groups. p-value: if a p-value < 0.05, it indicates there issignificant difference between the the dRMS of σ2s and IBS. . . . . . . . . . . . . . . 804.3 Average dERR of 36 samples in 9 subjects (4 samples per participant) at the suprail-iac, triceps and thigh sites respectively. The one-way ANOVA test is used to comparethe mean difference among results obtained from the SF, MF1 and MF2. F(2,105):F-value of the one-way ANOVA test with a between-groups degree of freedom of 2and a within-group degree of freedom of 105. p-value: if a p-value < 0.05, it indicatesthere is a significant difference among the groups. . . . . . . . . . . . . . . . . . . . . 82List of Tables viii4.4 Tukey’s HSD multiple comparisons for the difference in dERR within a group. SF:single focus at 25mm, MF1: stitching multiple focuses and MF2 averaging multiplefocuses. mean difference: the estimated statistical mean difference from the Tukey’sHSD test. CI : 95% confidence interval of the statistical mean difference betweengroups A and B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 824.5 Average dRMS of 36 samples in 9 subjects (4 samples per participant) at the suprail-iac, triceps and thigh sites respectively. The one-way ANOVA test is used to comparethe mean difference among results obtained from the SF, MF1 and MF2. F(2,105):F-value of the one-way ANOVA test with a between-groups degree of freedom of 2and a within-group degree of freedom of 105. p-value: if a p-value < 0.05, it indicatesthere is a significant difference among the groups. . . . . . . . . . . . . . . . . . . . . 844.6 Tukey’s HSD multiple comparisons for the difference in dRMS within a group. SF:single focus at 25mm, MF1: stitching multiple focuses and MF2 averaging multiplefocuses. mean difference: the estimated statistical mean difference from the Tukey’sHSD test. CI : 95% confidence interval of the statistical mean difference betweengroups A and B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 854.7 Summary on reference average thickness of subcutaneous fat collected from 9 par-ticipants with 4 samples for each person at each body site. The reference thicknessis obtained by manual segmentation on ultrasound data. . . . . . . . . . . . . . . . . 864.8 Discrepancies in skin-fold caliper measurement taken at the same spot of a body site. 944.9 The correlation coefficient r of average thickness between the manual ultrasoundsegmentation vs 12 skinfold caliper measurements for 18 samples in nine participants(two samples per participant). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954.10 The correlation coefficient r of average thickness between the automatic ultrasoundsegmentation vs 12 skinfold caliper measurements for 18 samples in nine participants(two samples per participant). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954.11 Mean difference and standard deviation values between the ultrasound measurementsand the 12 skinfold thicknesses. D is the mean difference and s is the standard deviation. 99List of Tables ix4.12 A comparison of correlations between ultrasound measurements and skinfold mea-surements at the suprailiac, thigh and triceps sites in this and past studies. Theultrasound measurements are obtained by automatic segmentation in this study,and by manual segmentation in the above past studies. . . . . . . . . . . . . . . . . . 100xList of Figures1.1 Human Subcutaneous Fat in an ultrasound B-mode image. . . . . . . . . . . . . . . 71.2 A group of adjacent elements activated simultaneously to generates an ultrasoundpulse and receives reflected echoes. A scan line consists of a single reflected RF signalthat is normally amplified and rectified. A B-mode image consists of a collection ofscan lines which are taken independently and combined to form an image of pixels.The scan lines are shown parallel here, but may also spread radially in a fan forcurvilinear transducers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.3 The mechanism of focusing using phased timing. . . . . . . . . . . . . . . . . . . . . 131.4 The ultrasound beam steering generated by phased timing. . . . . . . . . . . . . . . 141.5 B-mode images showing subcutaneous fat at different body sites. The thickness ofthe fat layer is indicated with arrows. . . . . . . . . . . . . . . . . . . . . . . . . . . 212.1 The concept of spatial compounding for reducing speckle in B-mode images. Afteraveraging images taken from different angles, speckle patterns are reduced and theobject appears more homogeneous. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.2 Application of spatial compounding on a spectrum property. . . . . . . . . . . . . . . 302.3 Mapping the coordinates from the raw data space to the real space by geometry. A1is a spectrum property value at its raw data coordinates (x1, y1) in the parametricimage obtained from the steering angle θ. A2 is the corresponding spectrum propertyvalue of A1 at the real spatial coordinates (x2, y2). . . . . . . . . . . . . . . . . . . . 312.4 The slope dfcdyof the linear regression line is proportional to the attenuation rateof a specific phantom layer. The red line is fc and the green dot line is the linearregression line of fc. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33List of Figures xi2.5 Reduction of the standard deviation in estimating the slope m of an fc scan lineusing spatial compounding with different step sizes of steering angle θ and differentnumbers of angles. The data points are fitted by the function 1x. . . . . . . . . . . . 342.6 Improvement in the standard deviation of estimating the slope m of the fc scan lineusing neighbour averaging with varying number of neighbour scan lines. . . . . . . . 352.7 Improvement in spectrum properties at a human suprailiac site after using spatialcompounding. (a) B-mode image. (b)Spectrum Properties(fc, σ2s and IBS). Theleft most column shows the spectrum properties without spatial compounding. Thesecond and third columns show the compound spectrum properties using 5 and 11angles of step size 2◦ respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.8 Improvement in spectrum properties at a human triceps after using spatial com-pounding. (a) B-mode image. (b)Spectrum Properties(fc, σ2s and IBS). The leftmost column shows the spectrum properties without spatial compounding. Thesecond and third columns show the compound spectrum properties using 5 and 11angles of step size 2◦ respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.9 Improvement in spectrum properties at a human thigh after using spatial compound-ing. (a) B-mode image. (b)Spectrum Properties(fc, σ2s and IBS). The left mostcolumn shows the spectrum properties without spatial compounding. The secondand third columns show the compound spectrum properties using 5 and 11 anglesof step size 2◦ respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.10 Detection of threshold using Rosin’s thresholding method on (a) unimodal histogramand (b) bimodal histogram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422.11 Results of thresholding on spectrum properties captured at the human suprailiacsite. (a) B-mode image. (b) Spectrum properties’ images(1st row), their histograms(2nd row) and binary maps (3rd row). The vertical line in the histogram indicatesthe calculated threshold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442.12 Results of thresholding on spectrum properties captured at the human triceps site.(a) B-mode image. (b) Spectrum properties’ images(1st row), their histograms (2ndrow) and binary maps (3rd row). The vertical line in the histogram indicates thecalculated threshold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45List of Figures xii2.13 Results of thresholding on spectrum properties captured at the human thigh site.(a) B-mode image. (b) 2nd to 4th row: a spectrum property’s image, histogram andbinary map respectively. The vertical line in the histogram indicates the calculatedthreshold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462.14 An example illustrates the extraction of boundary candidates in (a) a binary mapof σ2s using Equation 2.8 and (b) a binary map of IBS using Equation 2.9. Awhite pixel represents the value of one and a black pixel represents a pixel of zero.Blue solid crosses denote the boundary candidates obtained in the vertical directionand green dotted crosses denote the boundary candidates obtained in the horizontaldirection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482.15 Results of extraction and detection of boundary candidates from the binary mapBMσ2s obtained from a human suprailiac site: (a) potential boundary candidates(red crosses) and (b) fat boundary candidates found by RANSAC (red crosses). . . . 522.16 Results of extraction and detection of boundary candidates from the binary mapBMIBS obtained from a human suprailiac site: (a) potential boundary candidates(red crosses) and (b) fat boundary candidates found by RANSAC (red crosses). . . . 532.17 Results of extraction and detection of boundary candidates from the binary mapBMσ2s obtained from a human triceps: (a) potential boundary candidates (redcrosses) and (b) fat boundary candidates found by RANSAC (red crosses). . . . . . 532.18 Results of extraction and detection of boundary candidates from the binary mapBMIBS obtained from a human triceps: (a) potential boundary candidates (redcrosses) and (b) fat boundary candidates found by RANSAC (red crosses). . . . . . 542.19 Results of extraction and detection of boundary candidates from the binary mapBMσ2s obtained from a human thigh: (a) potential boundary candidates (red crosses)and (b) fat boundary candidates found by RANSAC (red crosses). . . . . . . . . . . 542.20 Results of extraction and detection of boundary candidates from the binary mapBMIBS obtained from a human thigh: (a) potential boundary candidates (redcrosses) and (b) fat boundary candidates found by RANSAC (red crosses). . . . . . 55List of Figures xiii2.21 Stitching of a spectrum property map obtained from multiple focuses: (a) stitchingof spectrum properties values (b) the weight function that combines two overlappingregions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562.22 The comparisons between (a)σ2s and (b)IBS obtained from: (1st-4th column) singlefocuses (SF) where F indicates the focus position, (5th column) stitching spectrumproperties from multiple focuses (MF1) and (6th column) averaging spectrum prop-erties from multiple focuses (MF2). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573.1 The framework of human subcutaneous fat detection. . . . . . . . . . . . . . . . . . 603.2 The sequence of capturing RF data. . . . . . . . . . . . . . . . . . . . . . . . . . . . 623.3 An example illustrates the segementation process on the parametric images of σsand IBS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.4 Body sites selected for skinfold caliper and ultrasound measurements. The directionof the arrows indicates the grasp of the skinfold caliper. . . . . . . . . . . . . . . . . 663.5 A Lange skinfold caliper. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673.6 A figure showing the mean difference D and the standard deviation of differences(s)between two methods. Example data are provided for illustration. . . . . . . . . . . 714.1 Examples (1st row: suprailiac, 2nd row: triceps and 3rd row: thigh) demonstrate thesegmentation results of σ2s at a single focus obtained from different structures andthicknesses of subcutaneous fat tissue. The cyan boundary is the manual segmenta-tion and the red boundary is the automatic segmentation. . . . . . . . . . . . . . . . 744.2 Two examples demonstrate the segmentation results of σ2s on participants with fatthickness ≤5mm. Ultrasound data is obtained using a single focus positioned at25mm. In sub-figures (a) and (b), the left image is the binary image of σ2s and theright image is the B-mode image. The cyan boundary is the manual segmentationand the red boundary is the automatic segmentation. . . . . . . . . . . . . . . . . . . 754.3 Correlation between manual and automatic measurements using σ2s at the (a) suprail-iac (b) triceps and (c) thigh sites. The red dashed line represents the one-to-onerelationship, the blue line is the linear regression line of the 36 samples (blue crosses). 77List of Figures xiv4.4 Correlation between manual and automatic measurements using IBS at the(a)suprailiac (b) triceps and (c) thigh sites. The red dashed line represents the one-to-one relationship, the blue line is the linear regression line of the 36 samples (bluecrosses). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 784.5 Show the results of IBS and σ2s at the (a) suprailiac, (b) thigh and (c) triceps sites.IBS is not reliable in locating the fat boundary at the thigh and triceps becauseof the presence of other soft tissues with strong reflection. In subfigures (a),(b) and(c), from left to right: binary map from IBS, segmentation result from IBS, binarymap from σ2s and segmentation result from σ2s . The cyan boundary is the manualsegmentation and the red boundary is the automatic segmentation. . . . . . . . . . . 874.6 An example showing the improvement of using MF2 over SF and MF1. In subfigures(a) SF,(b) MF1 and (c) MF2, from left to right: binary map of σ2s and, segmentationresult of σ2s . The cyan boundary is the manual segmentation and the red boundaryis the automatic segmentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 894.7 An example showing the improvement of using MF2 over SF and MF1. In subfigures(a) SF,(b) MF1 and (c) MF2, from left to right: binary map of σ2s and, segmentationresult of σ2s . The cyan boundary is the manual segmentation and the red boundaryis the automatic segmentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 904.8 An example showing the improvement of using MF2 over SF and MF1. In subfigures(a) SF,(b) MF1 and (c) MF2, from left to right: binary map of σ2s and segmentationresult of σ2s at a triceps. The cyan boundary is the manual segmentation and thered boundary is the automatic segmentation. . . . . . . . . . . . . . . . . . . . . . . 914.9 The relationship of the average thickness between the ultrasound and the skinfoldmeasurements at the suprailiac site: (a) manual ultrasound detection vs 12 skinfold(b) automatic ultrasound detection vs 12 skinfold. . . . . . . . . . . . . . . . . . . . 964.10 The relationship of the average thickness between the ultrasound and the skinfoldmeasurements at the triceps : (a) manual ultrasound detection vs 12 skinfold (b)automatic ultrasound detection vs 12 skinfold. . . . . . . . . . . . . . . . . . . . . . 97List of Figures xv4.11 The relationship of the average thickness between the ultrasound and the skinfoldmeasurements at the thigh: (a) manual ultrasound detection vs 12 skinfold (b) auto-matic ultrasound detection vs 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 985.1 Moving the transducer parallel to the skin to generate cross sectional images in avolume. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106C.1 Finding the threshold by Rosin’s thresholding method. . . . . . . . . . . . . . . . . . 120xviGlossaryIn vitro In an artificial environment outside the living organism.In vivo Within a living organism.A-mode The amplitude mode of ultrasound imaging.Adipocyte A fat cell from human tissue.Adipose tissue Also known as fat tissue. Is a layer of loose connective tissue specialized in storinglipids.Agar A gelatinous material that controls the stiffness of a phantom.Air-displacement plethysmography A device that measures the percentage of body fat byimmersing a subject in a closed air-filled chamber.ANOVA Analysis of variance.Attenuation Decrease in amplitude and intensity over a distance travelled by a wave.B-mode The brightness mode of ultrasound imaging.Beam focusing The concentration of the ultrasound beam into a small beam area.Beam steering The change of direction of the ultrasound beam.BIA Bioelectrical impedance analysis. A equipment that measures fat in terms of tissue conduc-tivity.Bimodal histograms A histogram with two distinct modes.Binary image An image consisting of only zeros and ones.Glossary xviiCellulose A complex carbohydrate that controls the scattering of a phantom.Collagen A long, fibrous protein structure that makes up the connective tissue.Connective Tissue A type of soft tissue that contains collagen. It connects and supports organsand tissues of the body.CT Computed tomography.DEXA Dual Energy X-ray Absorptiometry. A method that uses two low doses of X-ray beamswith different energy levels to detect bone and soft tissues.DFT Discrete Fourier transform.Echogenicity The relatively strength of echoes. A higher echogenicity (i.e. hyperechoic) means atissue structure having relatively strong echoes; a lower echogenicity (i.e. hypoechoic) meansa tissue structure having relatively weak echoes.Elastography The measurement of the elastic properties of tissue with ultrasound.Energy absorption Energy is absorbed and converted to heat when the sound wave propagatesthrough tissue.Fascia A sheet of fibrous connective tissue covers and separates muscles, organs, and other softtissues. It can be used to refer to the strong acoustic interface between subcutaneous fat andthe muscle layer.IBS Integrated backscatter coefficient. It indicates the total reflected power and strength of thebackscattering.Interface A surface that separates two kinds of soft tissues.Lagrange multiplier A method for optimizing a function that has several variables subject toone or more constraints.MF1 The method of stitching values of a spectrum property obtained at multiple focal positions.Glossary xviiiMF2 The method of averaging values of a spectrum property obtained at multiple focal positions.MRI Magnetic resonance imaging.Obese A person is obese if an excessive storage of body fat (women with more than 35% of bodyfat and men with more than 25% of body fat is accumulated in the body.)Parametric image A two-dimensional image of a spectrum property whose values are encodedinto gray-intensity scale.Phantom An artificial object that has some acoustic properties of soft tissues.Pulse-Echo Ultrasound Imaging A clinical ultrasound imaging technique that uses the sametransducer for both the pulse generator and echo detector.RANSAC Random Sample Consensus.Raw data coordinates The coordinates calculated with respect to the transducer.Reverberation Multiple ultrasound reflections between a structure and the probe.RF signal Radio-frequency signal is the raw signal of ultrasound.Sagittal The sagittal plane of the human body is an imaginary plane that symmetrically dividesthe body into right and left sections.Scan line A single line of ultrasound data that is parallel to the axial direction.Scattering The redirection of sound in several directions.Segmentation The partitioning of an image into two or more regions.SF Single focus at 25mm.Skinfold caliper A caliper that measures skinfold thicknesses by pinching a fold of skin and theunderlying subcutaneous fat.Spatial compounding The averaging of multiple overlapping data for reducing speckle noise.Glossary xixSpeckle Acoustic noise in ultrasound imaging due to the destructive and constructive interferenceof ultrasound pulses with randomly distributed scatters.Spectrum properties The properties of the power spectrum of the raw radio-frequency ultra-sound data.STFT Short time Fourier transform.Subcutaneous fat The fat layer that is immediately underneath the skin.Suprailiac The area on the side of human waist and on the iliac crest.TGC Time gain compensation.Thigh The area between the pelvis and buttocks, and the knee at the lower limb.Transducer A device, which consists of an array of piezoelectric elements, generates ultrasoundpulses by converting electrical energy to acoustical energy.Triceps A large three-headed skeletal muscle runs along the back of the upper arm of human.Underwater weighing A method that measures the percentage of body fat by immersing asubject in a water tank.Unimodal histograms A histogram with only one distinct mode.Visceral fat The fat layer that is located around internal organs.xxNotationBMIBS(x, y)) A binary image that contains the potential boundary candidates obtained fromIBS.BMσ2s (x, y) A binary image that contains the potential boundary candidates obtained fromσ2s .CI The confidence Interval.IBS(x, y, F ) A value of IBS at the coordinates (x,y) of the spatial compounded parametricimage obtained at the focal position of F .IBS(x, y, θ, F ) A value of IBS at the coordinates (x,y) of the parametric image obtained atthe steering angle θ and focal position of F .N The number of angles used in the spatial compounding.S(w) The power spectrum.W The bandwidth of the spectrum.dfcdyThe rate of change of central frequency along the depth.ˆIBS(x, y) The spatially compounded parametric image of IBS after combining valuesfrom different focal positions.σˆ2s(x, y) The spatially compounded parametric image of σ2s after combining values fromdifferent focal positions.λ The wavelength of an ultrasound pulseD The mean difference between two methods.φ The step size of angles used in the spatial compounding.σ2s Variance of the spectrum.σ2s(x, y, F ) A value of σ2s at the coordinates (x,y) of the spatial compounded parametricimage of obtained at the focal position at F .Notation xxiσ2s(x, y, θ, F ) A value of σ2s at the coordinates (x,y) of the parametric image obtained at thesteering angle θ and focal position of F .bBMIBS (x, y) A binary image that contains the fat boundary points obtained from IBS.bBMσ2s(x, y) A binary image that contains the fat boundary points obtained from σ2s .c The speed of sound.dERR The average thickness error of a detected boundary.dRMS The root mean square error of a detected boundary.fo The central frequency of the transducer.fc The central frequency of the power spectrum.mj The jth moment of a power spectrum.s The standard deviation of difference between two methods.x(t) A window of RF signal.IBS(x, y) The parametric image of IBS for segmentation.σ2s(x, y) The parametric image of σ2s for segmentation.k A parameter of the RANSAC algorithm. It is the number of iterations requiredfor the algorithm.n A parameter of the RANSAC algorithm. It is the smallest number of pointsrequired to fit the model.t A parameter of the RANSAC algorithm. It is the threshold (in pixel) requiredto determine if the data fit well.IBS Integrated backscatter coefficient.xxiiAcknowledgementsFirst of all, I would like to thank my supervisors, Dr. Peter Lawrence and Dr. Robert Rohling fortheir guidance, direction, advices and support for my thesis project. Their expertise in ultrasoundimaging and signal processing, and constructive feedbacks were necessary for the success of thisproject. Not only I have gained more research experience, but also learned how to deal with difficultsituations.Many thanks to all the participants of the experiment. I really appreciate their contribution,time and patience in completing the user study. This project would not be accomplished withouttheir volunteer. I would also like to thank Barry Legh, the senior instructor from the HumanKinetics Department of UBC, for teaching me the techniques of using the skinfold caliper and tolend me the Lange skinfold caliper for the experiment.In addition, I would like to thank all the coworkers in the Robotics and Control Lab for shar-ing knowledge and maintaining an excellent research environment. The assistance in using theresearch package to capture RF data from the software engineers in Ultrasonix are also gratefullyappreciated.Many thanks to my wonderful friends – especially to Angie, Vincci, Sheffield, Joyce, Sophia,Vivian, Simon, Horace and Aaron. Thank you for their listening, sharing, encouragement and allthe fun activities to balance out my life. I greatly value their friendships and appreciate their beliefin me. Also, thank Joyce and Raymond for providing comments for the thesis.Finally, I have been fortunate to have Jo in my life. I appreciate his understanding, patience, loveand his faith in me. He has also provided me with technical opinions. Thank you for accompanyingme to go through all the ups and downs and supporting me every step along the way.I am indebted to my family for encouraging me to pursue this study. Their continuous support,encouragement, patience and love are important for me to stay focus in my research.1Chapter 1Background and IntroductionEarly interest in measuring human body fat distribution can be dated back to 1921 when Matiegka[1] developed body fat predictive equations from subcutaneous fat skinfold thickness, body length,width and circumferences. Body fat analysis has been useful in assessing obesity to prevent healthrisk, monitoring athletes’ health status for giving appropriate nutritional counselings and moni-toring body shapes and weight for sports competition such as gymnastics and wrestling. Skinfoldcaliper measurement, body density weighing and bioelectrical impedance analysis have been popu-lar in assessing body fat. Recent advances in measuring body fat include the introduction of clinicalimaging techniques such as computed tomography, magnetic resonance imaging and ultrasound.These techniques produce images of human anatomy and provide a more accurate technique forresearchers. Among these imaging techniques, we are especially interested in ultrasound imagingbecause of its portability, safety and relatively low cost.Diagnostic ultrasound is a non-invasive, portable imaging device used mainly for clinical diagno-sis concerning organs and soft tissue. Recently, there has been a growing interest in assessing bodycomposition using ultrasound – in particular measuring thickness of subcutaneous fat. For exam-ple, researchers proposed to use ultrasound in order to overcome the drawbacks of compressibilityand elasticity in skinfold caliper measurements [2]. Other proposals include using it for measuringthe thickness of subcutaneous fat of obese persons [3] and on elderly people or at sites that are notconvenient for skinfold caliper measurement [4]. Perin et al. [5] also evaluated the occurrence ofnatural variations in thigh and abdominal subcutaneous fat thickness related to the phases of themenstrual cycle by using ultrasound measurements. In animals, ultrasound was used to predictintramuscular fat percentage at regions of interest by texture analysis in live swine [6]. Meanwhile,Abe et al. [7] attempted to calculate the subcutaneous fat volume by multiplying the fat thicknessobtained from B-mode ultrasound by the skin surface area. In wrestling, Saito et al. [8] measuredthe subcutaneous fat thickness of wrestlers at specific body sites and developed equations to predictChapter 1. Background and Introduction 2percentage of body fat. All of these methods have involved the manual analysis of the ultrasounddata to extract quantitative measurements.This thesis proposes to develop an automatic method to detect the subcutaneous fat thicknessusing ultrasound in vivo . In this introductory chapter, we will first survey the current techniquesof human body fat measurement and describe our motivation. After that, we will present theprinciples of ultrasound based on the pulse-echo technique. The properties of subcutaneous fatthickness and the difficulties in ultrasound detection are then discussed. Lastly, the thesis objectivesand organization are presented.1.1 Current Techniques of Human Body Fat MeasurementCurrent techniques that measure human body fat can be divided into the following categories: bodydensity weighing, bioelectrical impedance analysis (BIA), skinfold caliper, and imaging techniquessuch as dual energy X-ray absorptiometry, computed tomography, magnetic resonance imaging andultrasound imaging. Although they have the same goal of measuring body fat, their assumptionsare different. For example, the body density weighing method estimates the percentage of body fatbased on the body density whereas the BIA measures the fat tissues in terms of tissue conductivity.The skinfold caliper measures skinfold thicknesses at specific body sites and the percentage of fatis calculated based on these measurements. Compared with the above methods, the imagingtechniques are more direct in measuring body fat as they can image fat directly as soft tissues. Thecurrent techniques of measuring body fat are summarized in this section.1.1.1 Body Density WeighingUnderwater weighing and air-displacement plethysmography are two common methods that esti-mate the percentage of body fat based on body density [9]. The body density can be computedfrom the body volume and mass.1.1.1.1 Underwater WeighingThe underwater weighing technique requires the subject to be immersed in a tank of water whilefully exhaling. The calculation of the body density is based on Archimedes’s principle. ThisChapter 1. Background and Introduction 3principle states that the weight loss under water is directly proportional to the volume of waterdisplaced. The fat tissues are less dense than the bones and muscles; therefore, a person with ahigher percentage of fat makes the body lighter in water [10]. This method is time consuming andits equipment requires a lot of space. The results can be affected by the amount of air existing inthe subject’s lungs, changes in hydration and proportion of bone minerals. Moreover, the subjectshave to be fully immersed in water and this may cause discomfort.1.1.1.2 Air-displacement PlethysmographyThe air-displacement plethysmography requires a subject to immerse in a closed air-filled chamber.At a fixed temperature, the body volume of the subject can be directly measured by Boyle’s lawwhich states an inverse relationship between the pressure versus volume. This method does notrequire the subjects to immerse in water. Also, multiple readings can be recorded in a short periodof time. Therefore, the air-displacement plethysmorgraphy has begun to replace the underwaterweighing method [9]. Further, a good linear correlation of 0.94 is shown between between theunderwater weighing and air-displacement plethysmography [11]. Nevertheless, the accuracy ofthis method can be affected by changes in breathing pattern and movement of the subject.1.1.2 Bioelectrical Impedance AnalysisThe bioelectrical impedance analysis equipment measures fat in terms of tissue conductivity. Leantissue and water conduct electricity better than fat tissue; therefore, the measurement of the resis-tance to electrical current can be used to estimate the percentage of body fat. For the traditionalbioelectrical impedance analyzer (e.g. Tanita BIA scales), spot electrodes are placed on a person’sbare feet. The resistance of a small electrical signal is measured as it passes through the body. Bymodeling a body as a cylindrical conductor with its length proportional to the subject’s height,the impedance index can be calculated as the ratio of height square to body impedance [9]. Thismethod is easy to operate, inexpensive, portable and fast (less than 1 minute). However, it tends tooverestimate the body fat in obese people, and cannot distinguish between body fat and fluid[12].Chapter 1. Background and Introduction 41.1.3 Skinfold CaliperThe skinfold caliper measures skinfold thicknesses by pinching a fold of skin and the underlyingsubcutaneous fat. The average of multiple readings is needed at each body site to enhance accuracy.The common practice is to obtain skinfold thicknesses at three or four body sites and estimate thepercentage of body fat using predictive equations [13].The predictive equations calculate the percentage of body fat by substituting the values of fatthicknesses measured at several sites into a formula. Skinfold test formulas exist in many formsand are derived by human skinfold experiments. These formulas make use of the fat thicknessesmeasured at several sites for reducing measurement errors to calculate body fat percentage. Forexample, the equations developed by Jackson and Pollock [14, 15] used fat thicknesses measurementsfrom the triceps, suprailiac and thigh sites for females, and the chest, abdominal and thigh sitesfor males. Yuhasz [16] used the fat thickness measured at the triceps, subscapular, supraspinal,abdominal, thigh and calf sites for all male and female subjects. The constants in the equationscan be different between males and females.This method is the most widely used tool for evaluating body fat as it is fast, inexpensive andconvenient [17]. However, there are drawbacks of this method. The precision of the measurementscan be affected by the compressibility, thickness and water content of the subcutaneous fat layer,and also the elasticity of skin. Therefore, it is not possible to make precise measurements in obesepeople, the elderly, athletes in training, and those experiencing rapid weight gain or loss [2, 13]. Thequality of the calipers is also a factor: skinfold calipers should be accurately calibrated and shouldhave a constant specified pressure applied. Also, the precision of the method heavily depends onthe skill of a technician. Furthermore, the skinfold caliper is not suitable for all body locations. Forinstance, Nordander et al. [4] attempted to measure skinfold thickness directly over the trapeziusbut was not successful. The failure was due to the difficulty of grasping the skinfold.1.1.4 Imaging TechniquesDual Energy X-ray Absorptiometry (DEXA), Computed tomography (CT), magnetic resonanceimaging (MRI) and ultrasound are commonly used imaging techniques for clinical diagnostic pur-poses and they have also been introduced to quantify human body fat.Chapter 1. Background and Introduction 51.1.4.1 Dual Energy X-ray AbsorptiometryThe Dual Energy X-ray Absorptiometry uses two low doses of X-ray beams with different energylevels to detect bone and soft tissues. By assuming constant attenuation of the pure fat and leantissues within the soft tissues, the portion of fat and lean can be interpolated from each soft tissuepixel [18]. Tothill et al. [19] show a 15% difference in body composition was noticed between theequipment produced from different manufacturers. The accuracy of DEXA is dependent on thetechnology, method of calibration and interpolation of fat tissue. DEXA is costly and not portable.It also exposes subjects to ionizing radiation hazards. A trained technician is required to operatethe equipment. In addition, DEXA provides projection images and can only present the percentageof body fat that represents a substantial region.1.1.4.2 Computed TomographyCT is a radiological technique that generates cross section images of human anatomy using X-raybeams. By measuring the intensity of attenuated X-ray beams, the fat tissue, lean tissue andbones can be recognized. The fat tissue can be recognized as attenuation values between -190 to 30Hounsfield units (HU). Although slightly different intervals of attenuation values can be observedbetween investigators, they only have minor influence on the results [20]. Since the 1980s, severalstudies have measured the areas of abdominal subcutaneous and visceral fat using CT [21] andcalculated the fat tissue volume [22, 23] . The volume of fat could be calculated by multiplyinga cross-sectional area of fat tissue by the distance between each slice [22]. CT is used as a goldstandard of body fat measurements because of its excellent accuracy and precision [17, 25]; however,its immobility, high cost and exposure to a high dose of radiation make it inappropriate for frequentuse.1.1.4.3 Magnetic Resonance ImagingMRI is an imaging technique that uses both a strong magnetic field and a radio frequency electromagnetic pulse. The nuclei of hydrogen atoms in human soft tissues interact with the magnetic fieldthat is generated from the machine. Then, the pulsed radio frequency is applied to interact withthe hydrogen protons. After that, the radio frequency pulse is turned off and the protons releasethe absorbed energy at a certain rate. The rate of energy release, which is the relaxation time, isChapter 1. Background and Introduction 6related to the properties of the soft tissues. The adipose tissue (i.e. fat tissue) has a typical, shortlongitudinal relaxation time compared to other tissue [26]. High contrast between the fat tissueand adjacent muscles can be generated by a T1 weighted inversion recovery pulse sequence [27].Foster el al. [27] first applied MRI to research in body composition in 1984. MRI was used tocharacterize the distribution of human subcutaneous fat tissues in 1988 [28]. The fat and lean tissueswere observed at the mid-abdomen level and also for the whole body [29]. As in the case of CT,the volume of fat can be calculated by multiplying the cross-sectional area by the thickness of eachimage slice. Despres [20] summarized that MRI has a higher expected error for the measurementsof visceral fat than the measurements of subcutaneous fat. The coefficient of variation is in therange from 1.1% to 10.1% for the repeated measurements of subcutaneous fat and is in the rangefrom 5.3% to 10.6% from the repeated measurements of visceral fat. He also concluded that CTand MRI are the methods of choice for precise measurement of the subcutaneous and visceral fat.In terms of segmentation, Positano et al. [30] investigated the unsupervised segmentationof both abdominal subcutaneous and visceral fat tissue by fuzzy clustering approach using MRIimages. High linear correlations were found in the segmentation of both subcutaneous fat (r =0.9917) and visceral fat (r = 0.9601) when the results are compared with manual segmentation.Their method of segmentation overestimated the volume of subcutaneous fat with a mean per-centage difference of 6.4%, but underestimated the volume of visceral fat by a mean percentagedifferent of 7.9%.1.1.4.4 UltrasoundUltrasound is sound at frequencies that are above the range of human hearing: from 20kHz to severalhundred MHz. A higher frequency of ultrasound gives a better resolution, but, in turn, has a lowerpenetration power. Medical ultrasound usually uses frequencies from 1MHz to 10MHz; however,high frequency ultrasound ranging from 20MHz to 45MHz has also been used in characterizingrelatively shallow skin structures. Tissue boundaries can be distinguished because ultrasound pulsesare reflected at interfaces between tissues with different acoustic properties. The amplitude mode(A-mode) and the brightness mode(B-mode) are two common modes of displaying the reflectedultrasound echoes of soft tissue. Both of the modes involve the use of a focused ultrasound beamthat interrogates tissue along a line in space. In A-mode, reflected echoes are represented asChapter 1. Background and Introduction 7Figure 1.1: Human Subcutaneous Fat in an ultrasound B-mode image.amplitude versus depth in one-dimension. In B-mode, multiple equally spaced beams are used andreflected echoes are represented as two-dimensional brightnesses images whose axes correspond tothe lateral and axial direction of the scanning plane. In this way, each column of a B-mode imagecan be considered as a form of A-mode data. Tissues with stronger reflection are represented bybrighter intensities in both A-mode and B-mode images.The first mention of measuring subcutaneous fat thickness with ultrasound was in 1966 [31]. Itwas not until 1984 that Volz and Ostrove [32] used a portable A-mode ultrasound to quantitativelydetermine subcutaneous fat thickness in college woman. A-mode ultrasound measures the fatthickness by estimating the time required by an ultrasound pulse to be reflected from the fascia(i.e. the strong acoustic interface between subcutaneous fat and the muscle layer). They founda lack of agreement between ultrasound and half skinfold measurements: agreement ranged from87% at supraprailiac to 141% in thigh. The large range of error was believed to be caused bymultiple echoes reflected by the connective tissue layers dispersed within the fat tissues. Therefore,A-mode ultrasound alone is not sufficiently reliable to measure fat thickness. The later introductionChapter 1. Background and Introduction 8of B-mode ulrasound machines, which generate cross-sectional slices instead of just lines, makesit easier to interpret the depth of subcutaneous fat as two-dimensional ultrasound can depict thestructural information about the subcutaneous fat layer. Various studies have been carried out todemonstrate the reliability and reproducibility of B-mode ultrasound for quantifying subcutaneousfat. For instance, Kuczmerski et al. [3] measured subcutaneous fat thickness on obese adults toovercome the limitations of the skinfold calipers in 1987. They proved that ultrasound is superiorto the caliper technique in the prediction of the body density of obese persons. Bellisari et al. [33]recommended ultrasound as a measurement tool for subcutaneous fat. They evaluated the intraand inter-observer error and found that technical error was less than 0.2mm except in the femaletriceps where the inter-observer error was found to be 0.62mm. The reliability between observerswas above 90% except in the paraspinal site (82%). Flygare et al. [34] proved that ultrasound canreproduce measurements of subcutaneous fat in infants when performed by the same operator. In1996, Abe et al. [7] attempted to calculate subcutaneous fat volume by multiplying the fat thicknessobtained from B-mode ultrasound with the skin surface area. They found that the volume of fatmeasured by ultrasound was significantly correlated (r = 0.79-0.95) with MRI measurements atthe forearm, upper arm, trunk, thigh and lower leg. Other researchers using manual ultrasoundmethods have found that there are also high correlations between ultrasound and skinfold methods[32, 35, 36, 37, 2].CT and MRI are able to measure visceral fat by subtracting subcutaneous fat from the totalfat tissue[20]. However, ultrasound is not a good choice for quantifying visceral fat due to its lowresolution and poor penetration. Nevertheless, Tornaghi et al. [25] proposed one valid methodto indirectly assess the amount of visceral fat with ultrasound by measuring the intra-abdominaldepth that correlates to the amount of visceral fat area.1.2 MotivationSubcutaneous fat thickness is accepted as a body fat indicator because about 40 to 60% of total bodyfat is in the subcutaneous regions [38] and it is appropriate to use the distribution of subcutaneousfat as an indicator. Measuring subcutaneous fat is important in relation to human health problems,athletic performance and general public interest.Chapter 1. Background and Introduction 9Assessing human health risk – Body fat percentage can be used to assess obesity and humanhealth risk. A person is obese if an excessive storage of body fat (women with more than 35% of bodyfat and men with more than 25% of body fat [39]) is accumulated in the body. It is generally agreedby health professionals that an obese person has a higher chance of developing health problemssuch as hypertension, coronary artery diseases, stroke, gallbladder diseases, osteoarthritis, type 2diabetes, sleep apnea, respiratory problems, and cancers. Obese individuals may also suffer fromsocial stigmatization and discrimination [40].Evaluating performance of athletes – For professional athletes, monitoring their bodyweight and percentage of body fat can help to improve and maintain their sports performance.Body weight cannot be the sole indicator relating to the performance of athletes: for the sameweight of bodies, more strength and endurance can be generated from bodies composed of moremuscle than fat. Therefore, an optimal percentage of body fat can help the athlete to achieve thedesired performance in speed, agility, strength and endurance. The optimal percentage of body fatdepends on the nature of the sport, the sex of the athlete, and is determined on an individual basis.For example, gymnasts and figure skaters maintain little body fat for appearance and agility whileachieving optimal strength. Power sports such as football, skiing, volleyball and hockey requiremore fat (5 to 19% in males and 10 to 20% in females [41, 42]) to achieve a higher strength-to-weightratio for generating power. Sports like wrestling, weight lifting and body building set limitations interms of body weight. Saito et al. [8] state that it is important to develop a method for measuringthe body fat percentage of sumo wrestlers so that their weight can be monitored to prevent obesity-related diseases and also to maintain their competitive athletic performance and qualification for aspecific weight category. Athletes may suffer from eating disorders and poor energy level when thelevel of body fat is too low and this can be an indication of over-training. There is also a trend infemale atheletes to suffer from disruptions of menstrual cycles [43], amenorrhea and osteoporosis[42].General public interest in modern culture – Our modern culture and fashion industrytend to emphasize personal body shape, appearance and slimness. Under peer pressure, peoplemay consider being fat or carrying excessive fat at a particular body site to be unattractive. Obeseindividuals may also be afraid of discrimination. Due to these social and psychological issues,people are eager to monitor their body fat distribution, not only to minimize health problems, butChapter 1. Background and Introduction 10also for the sake of appearance.In general clinical practice, medical practitioners commonly use BMI as an indicator of obesitybecause it is easy to obtain. However, this method is not reliable [44]. On the other hand, itis inappropriate to assess athletes based solely on body weight because this does not reflect thepercentage of body fat. Development of a precise method to measure subcutaneous fat will allowpeople to monitor their body fat percentage. Based on the measurements, health care practitionerscan provide diet and nutritional counseling, recommendations on aerobic and exercise activitiesthat will prevent and control obesity, get the body in shape and set personal fitness goals.As discussed in section 1.1, there are several methods to estimate body fat percentage. Com-pared with techniques like body density weighing, BIA or skinfold caliper measurement, ultrasoundimaging is superior because it can provide real images of fat. Moreover, ultrasound imaging is moreportable and cheaper than DEXA, CT and MRI. Also, it does not require a radiation dose. Theabove factors make it ideal for measuring subcutaneous fat thickness for general use. Moreover,ultrasound imaging can be used to observe the macroscopic structure of the fat layer [2]. Studiesalso show that ultrasound is reliable and its results are repeatable for inter-observer data. [33, 45].The above reasons have motivated us to investigate the possibility of the automatic detectionof human subcutaneous fat. Although ultrasound has been introduced to measure subcutaneousfat for a few decades, we are not aware of any published work related to the automatic detectionof subcutaneous fat in vivo. Indeed, Glasbey et al. [46] investigated the automatic interpretationof subcutaneous fat in sheep using B-mode imaging. However, their automatic method only in-terpreted fat boundaries at two locations: the last rib and the third lumbar vertebra where theanatomy is relatively simple and presumed the skin boundary was known. Automatic detectioncan overcome the tedious task of manual detection, reduce the discrepancy among judgments ofdifferent operators, and standardize the measurement technique. We believe that automatic de-tection of fat thickness using ultrasound may potentially help to bring the technology into generalapplication by making it more user-friendly.Chapter 1. Background and Introduction 111.3 Pulse-Echo Ultrasound ImagingUltrasound machines, which are mainly used for medical applications, are based on the pulse-echo technique. In other words, the same transducer acts as both the pulse generator and echodetector. Ultrasound is generated in pulses and these pulses are reflected from different body softtissues as the pulse propagates through tissues. These echoes are separated in time and the timeis in proportion to the depth of the tissue interfaces. The depth and location of soft tissues canbe calculated by the arrival time of a reflected echo using the generalized speed of sound of softtissues. Moreover, the strength of the reflected echo indicates the difference in acoustic properties atthe interfaces. This section describes how an ultrasound machine works, how ultrasound interactswith soft tissue, the influence on the radiofrequency (RF) spectra and the current development ofultrasound image segmentation.1.3.1 ApparatusAn ultrasound machine consists of three main components: a transducer, an image processingunit and the display unit. The transducer consists of an array of piezoelectric elements whichgenerates ultrasound pulses and receives reflected echoes. A group of adjacent piezoelectric elementscan be activated simultaneously generates an ultrasound pulse by converting electrical energy toacoustical energy. The same group also receives reflected echoes as electrical voltages. The reflected,unprocessed radio-frequency echoes are referred to as RF signals. A scan line of RF signal providesamplitude data of the reflected echoes in the axial direction and thus forms A-mode data. Alateral collection of RF scan lines provides the amplitude information in both the axial and lateraldirections and thus forms a B-mode image (Figure 1.2). Furthermore, the time sequence of firingof elements within the group can be controlled for the application of beam focusing and steering.Overall, the transducer plays an important role in beam focusing, beam steering and controllingthe penetration and resolution of an ultrasound pulse.Focusing sound beams can improve resolution by reducing the beam width. The ultrasoundtransducer can control the position of the focal point in three ways: using curved piezoelectricelements, a lens or by phased timing. The first two methods use the geometric shape of theelements or lens to reduce the ultrasound beam width. The phased timing method is more commonChapter 1. Background and Introduction 12TFigure 1.2: A group of adjacent elements activated simultaneously to generates an ultrasound pulseand receives reflected echoes. A scan line consists of a single reflected RF signal that is normallyamplified and rectified. A B-mode image consists of a collection of scan lines which are takenindependently and combined to form an image of pixels. The scan lines are shown parallel here,but may also spread radially in a fan for curvilinear transducers.in clinical ultrasound machines and allows the focus to be adjusted electronically. Figure 1.3 showsthat pulses can be fired by the element groups at separated time intervals to adjust the apparentcurvature of the beam wavefront for controlling the location of the focus. A longer delay betweenelements in Figure 1.3(a) increases the curvature of the beam wavefront and the focus is movedcloser to the transducer, whereas a shorter delay in Figure 1.3(b) decreases the curvature of thebeam wavefront and the focus is moved farther from the transducer. This technique allows the useof multiple pulses at each scan line for positioning more than one focal point at different depths;however, the frame rate is decreased due to the need to fire multiple pulses[47].Moreover, phase timing can also control the beam steering angle by introducing additionalphase delays between the firing of individual elements. Figure 1.4 shows that the application ofvoltage pulses on the element groups with no delays generates a steering angle of zero degrees(no steering), whereas applying voltage pulses with increasing delay from left to right changes theChapter 1. Background and Introduction 13(a) (b)Figure 1.3: The mechanism of focusing using phased timing.steering direction to the left and applying voltage pulses with increasing delay from right to leftchanges the steering direction to the right. Beam steering allows the anatomy to be viewed fromdifferent angles. This approach is especially useful for techniques that aim to improve image qualityby averaging multiple views of the anatomy and for techniques that generate panoramic images bystitching images together.Ultrasound focusing can improve resolution at the focal point by reducing the beam width,but the overall system resolution of an ultrasound imaging system is determined mainly by thefrequency of the ultrasound pulses. There is always a tradeoff between the system resolution andthe penetration power of ultrasound pulses. The penetration and resolution of an ultrasound pulseis directly related to the frequency (f) and the wavelength (λ) of an ultrasound pulse. f and λ arerelated by the speed of ultrasound (c) viac = fλ. (1.1)In most ultrasound machines, c is assumed to be 1540ms−1 because ultrasound travels at nearly thesame speed for most biological tissues. The higher the frequency, the lower the penetration abilityChapter 1. Background and Introduction 14timeDirection of BeamSteering AngleFigure 1.4: The ultrasound beam steering generated by phased timing.of the ultrasound is. The smaller the wavelength and the shorter the ultrasound pulse, the betterthe resolution is. Therefore, an increase in frequency reduces the penetration power but improvesthe resolution; a decrease in frequency increases the penetration power but reduces the resolution.The change of frequency can be achieved by changing the natural frequency of the piezoelectricelements.To display ultrasound echoes, the received echoes are first amplified by the time gain compen-sation (TGC). TGC normally increases with depth so that it compensates for the effect of tissueattenuation with depth. After TGC, envelope detection is applied to each scan line of RF dataand a brightness image is obtained. The resulting two dimensional brightness image is referred toas an ultrasound B-mode image. In order to display the image, the B-mode image then undergoesthe scan conversion process. In this process, the scanlines are mapped to real spatial coordinatesaccording to the dimensions and geometry of the transducer. After that, the amplitude of thesignals are calculated and logarithmically compressed to match the dynamic range of the monitor.Chapter 1. Background and Introduction 151.3.2 Interaction of Ultrasound with Soft TissuesWhen an ultrasound pulse is fired from the transducer and propagates through soft tissue, atten-uation of the ultrasound energy occurs. Absorption, reflection, refraction and scattering are themost important mechanisms in the attenuation of the ultrasound energy and will be discussed. Tosimplify the interaction process, we assume there are no diffraction effects from the transducer be-cause it is not directly related to the tissue properties. Moreover, the characteristics of attenuationare determined solely by the acoustic properties of soft tissue.Reflection occurs at a boundary between two tissues with different acoustic impedances (Z).The acoustic impedance of soft tissue is defined asZ = ρc (1.2)where ρ is the density in g/m3 and c is the speed of sound in m/sec. c is assumed to be fixed at1540m/sec in the ultrasound machine. Given an incident wave perpendicular to a tissue boundary,the amount of wave energy transmitted across the surface is determined by the ratio of acousticimpedances of the two adjacent tissue. The ratio R of reflected intensity Ireflected and the incidentintensity Iincident isR =IreflectedIincident= (Z2 − Z1Z2 + Z1)2 (1.3)Lower values of R means more energy is transmitted through the interface.Refraction occurs at a tissue interface when there are different sound speeds of the two tissuesand the wave incidence angle is not 90◦. If an ultrasound wave travels from a soft tissue of speedv1 to a soft tissue with a speed of v2, the relationship between the angles of incidence θincident andrefraction θrefracted is given by Snell’s law:θincidentv1=θrefractedv2. (1.4)The variation in the speed of sound in soft tissues causes refraction and may resulted in incorrectpositions of tissues in the B-mode images.Scattering occurs when structures inside tissues are about the same size or smaller than thewavelength of ultrasound. Unlike specular reflection, scattering causes the sound beam to beChapter 1. Background and Introduction 16reflected in several directions and this reduces the echo strength. The factors affecting the scatteringproperties include: the size and number of point scatterers per unit volume within a tissue, the shapeand structures of scatters and the acoustic impedance differences at the scattering tissue interfaces.Energy lost due to scattering is small when it is compared to energy absorption[48, 49]. Thebackscattering of tissues is related to the presence of collagen structures, such as microvasculatureand elastin fibers[49]. For example, collagen fibers produces stronger scattering than blood cells[50].Scattering contributes to the local echogenicity of a tissue region; however, the destructiveand constructive interference of ultrasound pulses with randomly distributed scatters can generatespeckle that does not reflect the structure of the underlying tissues. Speckle patterns are randomand they usually appear when the scatterers are smaller than the resolution of an ultrasound pulse.They reduce the contrast resolution and degrade the details of the image. They can be found inboth RF data and B-mode images; the texture of the speckle patterns does not necessarily reflectthe structure of the corresponding tissue.Energy is absorbed and converted to heat when the sound wave propagates through tissue. Theamount of energy absorbed is dependent on the relaxation phenomena of the translational androtational vibration modes of the biological macromolecules [50]. For example, lung tissue consistsof air sacs and has a very high attenuation, whereas degassed water has low attenuation.Generally, the overall attenuation is due to both scattering and energy absorption. The overallenergy lost is expressed by the attenuation coefficient (µ) which indicates the energy lost in decibelsper centimeter of travel. It is generally assumed that the attenuation coefficient in human tissue islinearly proportional to ultrasound frequency except in blood (1.25dBcm−1), bone(1.7dBcm−1) orlungs(0.6dBcm−1) [51, 52, 53]. Note that an increase in ultrasound frequency causes an increasein attenuation. Moreover, Goss et al. [52, 53] summarized the ultrasound properties of variousmammalian tissues and showed that the acoustic property of the same type of tissue can vary due tomany factors such as the temperature, location of the sample, homogeneity of tissue, experimentaltechniques and in vitro or in vivo experiments. For example, the attenuation of fat is 10-15% higherat the room temperature than at a temperature of 37oC [52].It can be seen that the interaction between ultrasound and tissue is complicated but trends andmodels exist. The interaction affects both amplitude and spectrum properties of the tissue due tothe frequency dependence of the interaction. However, B-mode images, which are commonly usedChapter 1. Background and Introduction 17in traditional clinical analysis, are normally displayed after envelope detection so the spectrumcontent is lost. Spectrum properties can be an important description of soft tissues with differentacoustic properties; therefore, spectrum properties of ultrasound RF signal will be discussed in thefollowing section.1.3.3 Backscattered Radiofrequency SpectraThe previous section describes how ultrasound waves propagate through soft tissues in terms ofthe tissue properties. This section describes the interaction based on the spectrum of RF data. Asimplified model of the pulse-echo ultrasound interaction in the frequency domain at a particularlocation can be described asR(f) = P (f)B(f)A(f) (1.5)where R(f) is the spectrum of the received echo and P(f) is the spectrum of a transmitted pulse.The interaction of the transmitted pulse and the soft tissues is characterized by the spectrumbackscattering B(f) and spectrum of attenuation A(f) [54].Assuming the transmitted pulse is Gaussian in shape, P(f) can be described asP (f) = Poe−(f−fo)22σ2 (1.6)where Po is amplitude constant, fo is the transducer central frequency and σ is the bandwidth ofthe pulse spectrum. The backscattering spectrum can be modeled asB(f) = Bofz (1.7)where z is the scatter power and ranges from 0 to 4, and most human tissue is within the range ofz = 1 to 2 [49]. On the other hand, the attenuation spectrum can be modeled asA(f) = Aoe−αf (1.8)if we assume attenuation in human tissue is a linear function of frequency and ultrasound attenuatesexponentially [48]. α is the total accumulative round trip attenuation in tissue. The typicalvalue of α in soft tissue is 0.5dB(MHzcm)−1 [55] and this value varies among soft tissues. ForChapter 1. Background and Introduction 18example, striated muscle has a value of 1.30dB(MHzcm)−1, fat tissue at 37◦C has a value of0.61dB(MHzcm)−1 and blood has a value of 0.15dB(MHzcm)−1 [52].By combining Equations 1.6 to 1.8, the spectrum of the received echo R(f) becomesR(f) = PoBoAoe−(f−fo)22σ2 e−αffz. (1.9)Therefore, R(f) can be expressed in proportion to f , fo and z as:R(f) ∝ e−[f−(fo−ασ2)]22σ2 fz (1.10).According to the work of Treece et al. [54], fz can be expressed as an exponential and thescaling factors that are not related to frequency f are dropped. As a result, R(f) can be simplifiedand expressed in terms of a received Gaussian pulse as:R(f) ∝ e−(f−fc)22σ2 (1.11)whereσ = σ2f2of2o + zσ2(1.12)andfc = fo − ασ2 +zσ2fo. (1.13)σ and fc are the bandwidth and central frequency of the received pulse R(f) respectively.From equation 1.13, it is noticed that the central frequency fc of the received pulse spectrumshifts down as the round trip attenuation α increases. α increases when the travelling depth ofthe pulse increases; therefore, fc decreases as the depth increases. If z changes abruptly from onetissue to another tissue, there will be an abrupt change in the center frequency fc and a reductionof the bandwidth σ.Moreover, Fink et al. [56] used the short time Fourier transform (STFT) analysis and showedChapter 1. Background and Introduction 19that the downshifting rate of fc is proportional to the attenuation α asdfcdy= cασ2(y) (1.14)where c is the speed of sound in soft tissue, σ2(y) is the bandwidth of the received spectrum thatchanges with depth (y) and dfcdyis rate of change of fc.The above equations have shown that soft tissues with different acoustic properties would havedifferent influences on the central frequency and bandwidth of the received echo spectrum.1.3.4 Current Development of Ultrasound SegmentationMost research in ultrasound segmentation is concentrated on B-mode images. For instance, Akgul etal. [57] used a deformable contour to detect the tongue boundary in B-mode images. An automaticfuzzy multi-resolution-based algorithm was developed for cardiac left ventricular epicardial andendocardial boundary detection [58]. Madahushi and Metaxax [59] automatically found lesionmargins in ultrasound images by using both empirical domain knowledge used by radiologist and lowlevel image features. Low level image features include texture, intensity and directional gradients.These are just a few examples of a large body of literature. However as mentioned, B-modeimages are formed after envelope detection of RF signals, so spectral information that describesthe properties of soft tissue is lost.Not until recently have researchers started to investigate boundary detection using RF signals.In 1999, Hammoude [60] first attempted to detect edges based on abrupt changes in the centralfrequency due to the attenuation rate. However, his method failed to correctly identify the bound-ary due to erratic changes in ultrasound signal. Boukerroui at al [61] investigated a 3-dimensionaladaptive clustering segmentation method for in vivo echocardiographic 3D data based on gray-scaleintensity, two-dimensional texture features calculated envelope data and the local mean frequencyof the spectrum. In 2003, Dydenko et al. extracted the power of the signal, spectral-based au-toregressive parameters, and a velocity-based parameter to detect boundaries in echocardiographicimages by an adaptive smoothing algorithm. They also obtained promising results from performingboundary detection in cardiac sequences in vivo using the variance of velocity [62]. Davignon etal. used the integrated backscatter and mean central frequency to improve their multi-resolutionChapter 1. Background and Introduction 20Bayesian region-based algorithm that was based on envelope data. Their algorithm was tested onagar-gel phantom and proved that a multiparameteric approach could improve the segmentationresult [63]. These examples show that it is feasible to use spectrum information of the RF signalsfor the purpose of segmentation.1.4 Properties of Human Subcutaneous FatThe properties of human subcutaneous fat are presented from the biological and the ultrasoundviewpoints. The potential difficulties in detecting the fat layer in B-mode ultrasound images arediscussed.1.4.1 Biological characteristicsFat, which is also known as adipose tissue, is a layer of loose connective tissue specialized in storinglipids. Fat cells are held together by thin fibrous membranes of connective tissue. Connectivetissue in fat usually appears as thin and relatively sparse collagen fibers. Fat cells found in humans(mostly white adipose tissue) are also known as adipocytes and consist mainly of lipids (80% ofa fat cell) and can range up to 120µ in diameter [64]. They are spherical in shape when isolated.However, fat cells are usually packed to form a meshwork and become polyhedrical in shape. Thenumber of fat cells increases mainly during our infancy. An adult gains weight mainly because ofan increased accumulation of lipids in fat cells and not an increase in the number of fat cells [64].In the human body, fat can be divided into two types: subcutaneous fat and visceral fat. Thesubcutaneous fat layer is immediately underneath the skin and usually found in the thigh, waist,abdomen and buttocks. Visceral fat is usually located internally, around kidneys, at mesentery andretroperitoneal spaces etc [65]. However, in this thesis, we are mainly interested in the subcutaneousfat layer.1.4.2 Ultrasound characteristicsAs mentioned before, fat tissue consists mainly of lipids with sparse connective tissues. In ul-trasound B-mode image (as shown in Figure 1.1), human subcutaneous fat is separated from thenext soft tissue layer (usually the muscle) by a continuous white layer called fascia. This layerChapter 1. Background and Introduction 21(a) suprailiac (b) triceps (c) thighFigure 1.5: B-mode images showing subcutaneous fat at different body sites. The thickness of thefat layer is indicated with arrows.usually generates a strong reflection of the ultrasound pulse and is referred to as the fat boundary.Connective tissues are also dispersed within the fat tissue.Figures 1.5(a) to 1.5(c) show the appearance of the subcutaneous fat layer in B-mode images.In this study, when we refer to the subcutaneous fat layer, it starts with, and includes, the dermis ofthe skin, and ends at the continuous white layer which is the fascia. When the layer of subcutaneousfat is homogeneous, it usually appears to be less echogenic than other tissue layers such as muscles(Figure 1.5(a)). In addition, fibrous membranes of connective tissue, whose length, thickness anddensity vary between people and body sites, can be found within the layer of fat. As a result, thelayer of subcutaneous fat appears to be more echogenic in the presence of thicker connective tissue(Figures 1.5(b) and 1.5(c)). From our images, we noticed that subcutaneous fat at the suprailiacsite (figure 1.5(a)) is usually more homogeneous (with less or thinner connective tissue) than atthe triceps and thigh. At the triceps and thigh (Figure 1.5(b) and 1.5(c)), the layer of fat usuallyconsists of longer, thicker and denser connective tissues.Researchers have conducted studies to investigate the acoustic properties of human fat. Sum-maries from [52, 53] show that fat has a relatively low speed of sound (∼1480ms−1) compared withChapter 1. Background and Introduction 22other human soft tissue, while connective tissue (∼1613ms−1) has a relatively high speed of sound.Therefore, fat tissue is highly heterogeneous and may cause ultrasonic wavefront distortion [66].Moreover, previous research has also shown that different types of fat tissues have dissimilaracoustic properties. For example, Greenleaf and Bahin [67] investigated fat in the breast usingtransmissive ultrasound computerized tomography and found that an increase in the number ofcollagen in fat content would increase its speed of sound and its attenuation. Landini and Sarnelli[68] also found that the attenuation coefficient is lower for tissues with a large predominance of fatcells and this increases with more collagen fiber content.The usual practice of a sonographer in determining the subcutaneous fat thickness is to drawa vertical line from the surface of skin to the fascia [13]. There are several factors affecting thevisual interpretation of the fat boundary. For example, dense connective tissues may appear nearthe fascia and make the fat-muscle boundary less clear, and a long connective tissue layer mayalso be wrongly interpreted as the fascia. Moreover, it is harder to define the boundary betweensubcutaneous fat and muscle because of the smaller amounts of intermuscular fat tissues[13].Although the fascia and fibrous connective tissues are well imaged because of their specularcharacteristics, the amplitude of the received pulse can be affected by the angle of incidence of theultrasound beam. In our images, the transducer was kept vertical to the skin so that the angle ofincidence was near 90◦.1.4.3 Difficulties in Segmentation of Fat in Ultrasound ImagesFrom the observations in the previous section, we noticed that the variations in echogenicity, sizeand density of connective tissue among different people and body sites make it difficult to extractthe fat layer from B-mode images alone. Several authors have also reported that heterogeneity ofsubcutaneous adipose tissue was observed in ultrasound and X-rays images [31, 32]. Additionalstrong interfaces could appear near skin and intermediate membranes dispersed through the fattissues. Other researchers also show that the thickness and texture of fat affects the overall ap-pearance of B-mode images. Haberkorn et al. [69] investigated the influence of the subcutaneousfat layer on the diagnosis of B-mode images. They mentioned that the size of fat clusters mightchange the ultrasound wave length, and showed that an increase in thickness of fat caused darkerand lower contrast images. Pomaroli et al. [70] performed a histologic analysis and showed thatChapter 1. Background and Introduction 23fatty tissues with more connective tissues appeared to be more echogenic in B-mode than fattytissues with fewer connective tissues. Moreover, Hinkelman et al. [66] also mentioned that thicklayers of fat may cause poor B-mode image quality at the abdominal wall because they distortultrasound beams due to their scattering and absorption effects.1.5 Thesis Objectives and OrganizationConventional B-mode ultrasound images describe tissue structure only in terms of echogenicity andtexture. Based on our observed difficulties, fat tissue does not have a consistent description of bothtexture and brightness; therefore, image segmentation of subcutaneous fat is difficult when appliedto B-mode images. Conversely, RF signals retain the frequency, phase and amplitude information.Spectrum properties of the RF signal give additional information related to the acoustic propertiesof fat; thus, we will investigate the feasibility of measuring the subcutaneous fat by detecting thechanges of the spectrum from one layer to another. Moreover, the existence of speckle also affectsthe texture of ultrasound images. Ultrasound speckle adds noise to the RF raw data, reduces thecontrast resolution and weakens the detectability of soft tissue layers. Therefore, the existence ofspeckle also imposes further challenges on the segmentation problem.This thesis explores the use of the spectrum properties of RF signals to detect the boundary ofthe subcutaneous fat layer and presents an image processing and boundary detection frameworkto automatically detect the fat boundary in vivo. An experiment with nine human subjects isalso presented to validate the accuracy of the method by comparing our automatic measurementswith manual measurements. Furthermore, the correlation between ultrasound measurements andskinfold measurements is also investigated. The thesis is organized into four chapters as follows:Chapter 2 describes our method of image processing and boundary detection using the RFsignals. The calculation of spectrum properties is presented and spatial compounding is introducedto reduce the noisy spectrum measurements. The values of the spectrum properties are encoded asgray-scale parametric images for segmentation. A new segmentation technique on selected spectrumproperties using thresholding and boundary detection is also discussed. At the end, capturing RFdata from multiple focuses is proposed to overcome the drawbacks of using a single focus.Chapter 3 shows the overview of the entire human subcutaneous fat detection framework.Chapter 1. Background and Introduction 24The steps of the detection framework include RF data capture, calculation of spectrum properties,preprocessing of spectrum properties map and segmentation. Moreover, our method is tested at thesuprailiac, triceps and thigh sites of nine volunteers and is compared to the skinfold caliper method.The procedures of the user study in collecting skinfold caliper measurements and ultrasound dataare discussed.Chapter 4 evaluates the results of our fat boundary detection method. First, our results arecompared with the manual detection on B-mode images in terms of average thickness error andRMS error. We also investigate whether data obtained from multiple focuses will improve therobustness of our detection algorithm. Furthermore, relationships between skinfold caliper andultrasound measurements are presented in terms of their linear correlation and mean differences.Chapter 5 summarizes our work and presents the future work. It also describes the remainingissues of our detection method using ultrasound.25Chapter 2Method in Developing ImageProcessing and Boundary DetectionIn the previous chapter, we have shown that the acoustic properties of soft tissues can be relatedto the spectrum of the received radiofrequency (RF) signal. This chapter proposes the use ofunprocessed RF signals to develop an image processing and boundary detection algorithm to extractthe human subcutaneous fat layer. We discover that the characteristics of human subcutaneous fattissues can be described by the local spectrum properties of RF signals. The spectrum properties areencoded into gray-intensity images called parametric images. Then, we develop an image processingand boundary detection method based on our observations on the parametric images. Moreover, wediscover that the method of spatial compounding increases the signal to noise ratio and improvesthe detection of subcutaneous fat from their parametric images of the spectrum properties. Withthe spectrum properties established through spatial compounding, a thresholding and boundarydetection method is proposed to segment the subcutaneous fat layer and locate the fascia - the fatboundary. Finally, we also consider using spectrum properties obtained from multiple focuses toimprove the segmentation result.2.1 Processing of Radiofrequency SignalRF signals are stochastic (i.e. next state of the signal is partly but not fully determined bythe previous state of the signal) and their spectrum characteristics vary with time; therefore,their spectrum properties are estimated by local spectrum calculations. In each RF signal, localspectra are calculated by short time Fourier transform (STFT). The total energy, mean centralfrequency and spectrum variance are calculated for each local spectrum and their relationship withthe subcutaneous fat layer is investigated.Chapter 2. Method in Developing Image Processing and Boundary Detection 26The reasons for using these three spectrum properties are as follows:• The total energy reflects the strength of echoes. Since echoes reflected from the fascia arerelatively strong, the total energy can indicate the location of the fascia.• The mean central frequency shifts downward with depth when the pulse propagates. Its rateof change is proportional to the attenuation coefficient of the soft tissue and can potentiallyserve as a descriptor of soft tissues with different acoustic properties.• The spectrum variance provides the deviation of spectrum values from fc. It gives the un-certainty measurement for the estimation of the mean central frequency. It can potentiallydescribe the shape and bandwidth of the power spectrum which vary among different softtissues.2.1.1 Calculation of Spectrum PropertiesSTFT is performed on each RF scan line. A window, which consists of 32 samples, is shifted downin depth with a 50% overlapping of the previous window. Then, a Hamming window and discreteFourier transform (DFT) are applied to the window. M-point DFT is performed by Matlab’s fftfunction. The spectrum is obtained by:S(w) = |X(w) ∗X(w)| (2.1)w ∈ {0, fo, 2fo, ...(M − 1)fo}where fo is the sampling frequency of RF signals, X(w) is the DFT applied to a windowed RFsignal x(t) and S(w) is the power spectrum with a length of M/2. As mentioned in Section 1.3.3,the spectrum is assumed to have a Gaussian distribution.The three spectrum properties (total energy, mean central frequency and spectrum variance)of S(w) are calculated within a bandwidth W but not the whole spectrum. The purpose is toeliminate unwanted frequencies that are considered to be non-significant data. We use W from0.5MHz to 11MHz for data obtained from a 6.6MHz transducer. Since the transmission frequencyof the transducer is 6.6MHz and its bandwidth is small, it is reasonable to restrict the signal toChapter 2. Method in Developing Image Processing and Boundary Detection 27this range. A rule of thumb for the axial resolution of a 6.6MHz transducer is 0.35mm1. Thewindow size of 32 samples, which corresponds to 0.62mm, is around twice the axial resolution ofthe transducer.The first spectrum property discussed is the total energy of the spectrum, also referred to as theintegrated backscatter coefficient (IBS)[71, 63]. The IBS coefficient indicates the total reflectedpower and strength of the backscattering; a larger value corresponds to more energy reflected fromthe tissues. The IBS is calculated by:IBS =Wmax∑w=WminS(w) (2.2)where w is from Wmin to Wmax MHz, S(w) is the power spectrum with a length of M/2.The mean central frequency (fc) describes the spectrum central frequency which is the averageof the frequencies present in a window. Section 1.3.3 show that the rate of change of fc is directlyproportional to the attenuation and its value fluctuates when the ultrasound pulse propagatesthrough tissue and is reflected by tissues with different acoustic properties. fc can be calculatedusing the moment analysis. Moment analysis has been used by several authors [56, 63] to calculatelocal spectrum properties of RF signals. Fink et al. [56] proved that the moment analysis can beused calculate the central frequency and variance of the power spectrum. In moment analysis, thejth moment mj is calculated as the following:mj =Wmax∑w=WminwjS(w) (2.3)where S(w) is the amplitude of the spectrum and wj is the jth power of w. By using Equation(2.3), fc is expressed by the zeroth moment(m0) and first moment (m1) of the spectrum [56] asfc =m1m0(2.4)The last spectrum property is the spectrum variance(σ2s). It tells the deviation of spectrumvalues from fc within a bandwidth W; a larger value corresponds to a higher uncertainty for the1The axial resolution (for a three cycle pulse) = (temporal pulse length x 3) x c/2 = (c x 3)/(fo x 2)Chapter 2. Method in Developing Image Processing and Boundary Detection 28estimation of fc. To our knowledge, this spectrum property is rarely used in investigating tissueproperties in ultrasound. Bylund at al [73] discovered that the estimated spectrum variance waslow at the reverberation artifact locations[74]. According to Fink et al. [56] σ2s can be calculatedusing the first and second moment(m2) of the spectrum and fc given by:σ2s =m2m0− f2c . (2.5)The calculation of σ2s , as proved by Fink et al. [56], is shown in Appendix A.2.1.2 Noise Reduction using Spatial CompoundingStatistical fluctuations exist in backscattered RF signals due to speckle noise and heterogeneityin tissues and result in noisy spectrum properties. Spatial compounding has been used to reducespeckle and improve boundary continuity in B-mode images [75, 76, 77]. Recently, compoundimaging is also applied to reduce the variance of displacement estimations in elastography [78], toimprove temperature estimations due to the thermo-acoustic lens effect in high intensity focusedultrasound [79] and to reduce variance of attenuation measurements and enable coarse attenuationimaging [80]. In our work, compound imaging is used to improve the estimations of IBS, fc andσ2s .The concept of spatial compounding is shown in Figure 2.1 for B-mode images. The object(gray circle) appears inhomogeneous in the presence of speckle and the speckle pattern changesunder different viewing angles. The two beam-steered B-mode images at left and right (not takenfrom the 0◦ direction), are transformed and interpolated from the raw data coordinates to Cartesianspatial coordinates. The raw data coordinates are with respect to the transducer and the spatialcoordinates are in the real space. In the end, the resulting images in their spatial coordinatesare averaged to form a single compounded image. Since speckle patterns are random, their shapeand distribution changes with beam angle, so averaging images from different views can reducespeckle and the object appears more homogeneous. Our idea of applying spatial compounding tothe spectrum properties is similar to the above discussion except spatial compounding is applied toobtain the spectrum properties values (IBS, fc and σ2s) instead of the B-mode pixel values. Thedetails of implementation will be discussed next.Chapter 2. Method in Developing Image Processing and Boundary Detection 29RawdataCoordinatesReal spatialcoordinatesCompound ImageAverageBeam steeringangleFigure 2.1: The concept of spatial compounding for reducing speckle in B-mode images. Afteraveraging images taken from different angles, speckle patterns are reduced and the object appearsmore homogeneous.2.1.2.1 ImplementationSpatial compounding is applied to each spectrum property as shown in Figure 2.2. RF data arefirst obtained from N different steering angles by beam steering of the transducer. For each steeringangle θ, a spectrum property (IBS or fc or σ2s) is calculated from each RF scan line as describedin section 2.1.1. As a result, N two-dimensional spectrum properties , whose width is the numberof RF scan lines and height is the number of STFT windows, are computed for each spectrumproperty. The values of fc, σ2s and IBS may be presented as two-dimensional images. We will referto these two-dimensional images of the spectrum properties as “parametric images”throughout thethesis.Chapter 2. Method in Developing Image Processing and Boundary Detection 30Figure 2.2: Application of spatial compounding on a spectrum property.The next step is to convert a parametric image from its raw data coordinates to the real spatialcoordinates using the geometry of the steering (Figure 2.3).Chapter 2. Method in Developing Image Processing and Boundary Detection 31A1(x1,y1)A2(x2,y2)(0,0)xllφφyFigure 2.3: Mapping the coordinates from the raw data space to the real space by geometry. A1is a spectrum property value at its raw data coordinates (x1, y1) in the parametric image obtainedfrom the steering angle θ. A2 is the corresponding spectrum property value of A1 at the real spatialcoordinates (x2, y2).Let A1 be a spectrum property value at the raw data coordinates (x1, y1) in the parametricimage obtained from the steering angle θ and A2 be the corresponding spectrum property value ofA1 at the real spatial coordinates (x2, y2). Using geometry (Figure 2.3), (x2, y2) can be calculatedas:φ = −θℓ = y1x1 − x2ℓ= sinφx2 = x1 − ℓsinφ (2.6)y2ℓ= cosφy2 = ℓcosφ (2.7)Thereafter, a two-dimensional bilinear interpolation is applied to find the interpolated spectrumproperty value A2(x2, y2) in the real spatial coordinates space. The Matlab function interp2 isused for the the bilinear interpolation based on the known original raw data coordinates, real spatialChapter 2. Method in Developing Image Processing and Boundary Detection 32coordinates and values of the parametric image in the raw data coordinates space.2.1.2.2 ExperimentsA custom-made phantom is used to find the best range and step size of steering angles for spatialcompounding by calculating the uncertainty in estimating the rate of change of fc. Moreover, thespatial compounding method is compared to neighbour averaging to show that the reduction ofspeckle noise in spectrum properties is more effective when averaging from different viewing anglesinstead of simply from neighbouring data. The best range and step size of steering angles obtainedfrom the phantom experiment are then used for observing the qualitative improvements on theparametric images of IBS, fc and σ2s in real human tissue.2.1.2.2.1 Phantom Experiments Given a phantom that is homogeneous and composed ofonly one type of tissue with a constant attenuation rate, the rate (dfcdy) of down-shifting fc alongthe depth (y) direction is then linearly proportional to its attenuation rate. Since the value of fcvs y is erratic, we can characterize the change of fc for each RF scan line by linear regression asshown in Figure 2.4. (dfcdy) is then the slope of the linear regression line. The standard deviation ofslope values among all scan lines indicates the uncertainty in estimating dfcdyof a homogeneous layer.If the standard deviation is small, there is less statistical fluctuation of the spectrum value. Weused a homogeneous phantom to investigate the improvement in the standard deviation by usingspatial compounding and contrast the advantage of using spatial compounding over averaging withneighbouring values.2.1.2.2.1.1 Method A homogeneous phantom was constructed by agar and cellulose - agarcontrols the stiffness of the phantom while cellulose controls the scattering. The phantom consistsof 1% cellulose, 3% agar and 96% water. The transducer was mounted on a stand and stabilizedwith a clip during the experiment. We collected 127 scan lines of RF data from the phantom.RF data are used to compute fc by the method described in the Section 2.1.1 and are spatiallycompounded in accordance with the Section 2.1.2.1.The above procedure was tested with step size angles of 0.5◦, 1◦, 2.0◦ and 3.0◦. In the regionof interest, the linear regression fitting was applied at each scan line to compute the slope dfcdy. Thestandard deviation of dfcdybetween scan lines was calculated.Chapter 2. Method in Developing Image Processing and Boundary Detection 33DepthyfcSlope =Figure 2.4: The slope dfcdyof the linear regression line is proportional to the attenuation rate of aspecific phantom layer. The red line is fc and the green dot line is the linear regression line of fc.In the alternative method of neighbour averaging, values of fc are averaged with values fromneighbour scan lines. A set of fc values along a scan line is averaged with values from n scanlines(from both left and right neighbouring lines).The slope dfcdycan again be computed for each scanline of fc. The standard deviation ofdfcdybetween scan lines is also calculated.2.1.2.2.1.2 Results and Discussions Figure 2.5 shows the normalized standard deviation ofthe regression line slope dfcdyversus the number of angles N used for a given angle step size φ. Thestandard deviation is normalized to that obtained without spatial compounding or neighbor averag-ing. Figure 2.5 shows that spatial compounding helps to reduce the uncertainty in measurements.For a particular angle step size φ, the standard deviation decreases when the number of angles Nused in spatial compounding increases. This indicates that greater numbers of frames increases thesignal to noise ratio. In addition, more improvement is noticed when using a step size of 2.0◦ and3.0◦ than 0.5◦. It is because a larger separation of angle produces greater independence of the RFlines, so more speckle is reduced by averaging.Spatial compounding shows a more convincing improvement in reducing noise in spectrum prop-erties than neighbour averaging. Figure 2.6 shows the normalized standard deviation of regressionline’s slope versus the number of neigbouring scan lines used for averaging spectrum parameter val-ues. The standard deviation is normalized by the value without neighbour averaging. The resultshows that neighbour averaging does not produce a significant decrease in the standard deviationChapter 2. Method in Developing Image Processing and Boundary Detection 34as the number of neighbouring lines increases and it does not improve the standard deviation toless than 0.9. Therefore, neighbour averaging does not improve the signal to noise ratio as muchas spatial compounding because the data come from neighbouring lines are more correlated thandata from different angles.StepsizeNumber of angles (N)NormalizedSDFigure 2.5: Reduction of the standard deviation in estimating the slope m of an fc scan line usingspatial compounding with different step sizes of steering angle θ and different numbers of angles.The data points are fitted by the function 1x.Chapter 2. Method in Developing Image Processing and Boundary Detection 35Figure 2.6: Improvement in the standard deviation of estimating the slope m of the fc scan lineusing neighbour averaging with varying number of neighbour scan lines.Nevertheless, there are tradeoffs of using spatial compounding. The increase of the range ofangles improves the overall compounding effect, but the area covered by full compounding reducesas depth increases. We want to perform segmentation on a parametric image with every pixelcompounded with the same number of angles; therefore, we want to choose a step size angle φand a number of angles N that covers a fair amount of area and has a substantial noise reductioneffect. According to our experiments, a combination of N = 11 and φ = 2◦ is best. For φ = 2◦, thecompounding effect is similar to that of N = 9 and φ = 3◦. The former combination requires anangle range of (±10◦) and latter combination requires a larger range of angles (±12◦). In addition,if the steering angle is too large, the probability of echoes returned from tissue to the transducerbecomes lower. Based on the above reasons, we used N = 11 and φ = 2◦ (0◦,±2◦,±4◦,±6◦,±8◦and ±10◦) for our experiments with human in vivo.2.1.2.2.2 Human Experiments2.1.2.2.2.1 Method From the experiments with the phantom, it was shown that spatial com-pounding could improve the signal to noise ratio of spectrum calculation and the steering anglecombination of 0◦,±2◦,±4◦,±6◦,±8◦ and ±10◦ was chosen. A similar set of calculations was per-formed on the suprailiac, thigh and triceps sites of a human subject in vivo. Parametric imagesshown in the following sections are cropped to show the area with the same number of compoundframes N .Chapter 2. Method in Developing Image Processing and Boundary Detection 362.1.2.2.2.2 Results and Discussions Figures 2.7 to 2.9 show the normalized spectrum prop-erties with and without spatial compounding. Obvious quality improvements are shown in the fc,σ2s and IBS parametric images as the number of compounding frames increases. The paramet-ric images of fc and σ2s obtained from one angle (0◦) are erratic. Lack of uniformity is especiallyobserved at the triceps (Figure 2.8) and the thigh (Figure 2.9). At the above two sites, the subcuta-neous fat is less homogeneous than at the suprailiac site (Figure 2.7). The estimation of the powerspectrum is affected by speckle noise, but the spatial compounding reduces speckle noise in para-metric images of fc, σ2s and IBS by averaging their values obtained from different angles. Valuesare smoother and less erratic after spatial compounding; the characteristics of fat is differentiatedfrom other tissues after spatial compounding.The parametric images of both fc and σ2s show that their values decrease when depth increases,and their values are relatively higher in the fat tissue than other tissues. The discontinuity of theirvalues is also observed at the fat boundary. It is also observed that the structure of subcutaneousfat affects the appearance of fc and σ2s . For example, the B-mode image of Figure 2.8 shows thatthe subcutaneous fat is less homogeneous in the triceps when compared to that of the suprailiacsite(Figure 2.7). Dense fibrous membranes of connective tissue are also observed in the middle ofthe fat layer at the triceps. At regions near the connective tissue, there are sudden changes ofvalues in fc and σ2s . The fibrous membranes are the source of inhomogeneity in the fat tissue andthe cause of abrupt changes in the spectrum.IBS shows that strong interfaces happen at the fascia (the fat boundary) and fibrous membranesof connective tissue in the fat tissue. In addition, stronger interfaces also appear at the bone ortendon at the triceps and thigh. Parametric images of IBS show that these strong interfaces aremore complete after spatial compounding. Since strong interfaces are mostly smooth and specularsurfaces, viewing them at different angles increases the chance of achieving an incidence angle of 90◦when reflection is maximum. As a result, their boundaries appear more continuous and completein compound IBS parametric images.In general, fc, σ2s and IBS do show characteristics of subcutaneous fat. As observed, thesubcutaneous fat layers have relatively higher values of σ2s and fc than other tissues. However,higher contrast between the subcutaneous fat layer and non-fat tissue is noticed in the parametricimages of σ2s than in the parametric images of fc; therefore σ2s is more suitable and easier forChapter 2. Method in Developing Image Processing and Boundary Detection 37segmentation. In addition, IBS shows valuable information about strong reflected interfaces thattells the locations of the fascia (i.e. the location of the fat boundary) and the thick fibrous connectivetissue in the subcutaneous fat.Chapter 2. Method in Developing Image Processing and Boundary Detection 38(a)(b)Figure 2.7: Improvement in spectrum properties at a human suprailiac site after using spatialcompounding. (a) B-mode image. (b)Spectrum Properties(fc, σ2s and IBS). The left most columnshows the spectrum properties without spatial compounding. The second and third columns showthe compound spectrum properties using 5 and 11 angles of step size 2◦ respectively.Chapter 2. Method in Developing Image Processing and Boundary Detection 39(a)fcIBS(b)Figure 2.8: Improvement in spectrum properties at a human triceps after using spatial compound-ing. (a) B-mode image. (b)Spectrum Properties(fc, σ2s and IBS). The left most column showsthe spectrum properties without spatial compounding. The second and third columns show thecompound spectrum properties using 5 and 11 angles of step size 2◦ respectively.Chapter 2. Method in Developing Image Processing and Boundary Detection 40(a)fcIBS(b)Figure 2.9: Improvement in spectrum properties at a human thigh after using spatial compound-ing. (a) B-mode image. (b)Spectrum Properties(fc, σ2s and IBS). The left most column showsthe spectrum properties without spatial compounding. The second and third columns show thecompound spectrum properties using 5 and 11 angles of step size 2◦ respectively.2.2 Thresholding on Spectrum PropertiesBased on the visual observation of compounded spectrum properties, thresholding is proposed toseparate the subcutaneous fat layer and the non-fat layer. In a typical bilevel thresholding case,there should be two distinct modes in the parametric image histogram (i.e. a bimodal histograms);however, there can be cases when only one obvious peak or a very small second peak is found inthe parametric image histogram. Classic thresholding algorithms like Otsu’s method [81] assumesChapter 2. Method in Developing Image Processing and Boundary Detection 41the histogram is bimodal and obtains the threshold by minimizing the within-group variance. Forunimodal histograms, Rosin [82] developed a unimodal thresholding technique based on findinga corner in the histogram. His method can calculate a threshold in unimodal histograms and isnot affected by the distribution of histogram. Good results have been shown in various threshold-ing tasks [82] and remote sensing image thresholding [83] using his thresholding method. In ourthresholding task, we noticed that the intensity histogram of spectral parameters may be unimodal.Based on our assumptions of segmentation and observations, Rosin’s thresholding method is usedand will be discussed in this section.2.2.1 Unimodal ThresholdingTo use Rosin’s thresholding method, assumptions are made on the parametric images of the spec-trum properties:1. The subcutaneous fat (with skin) is always the first top layer of tissue; therefore, it has arelatively higher value of fc and σs than the values of the following layer if no noise is present.2. In the parametric images of IBS, strongly reflective tissues such as the fascia and fibrousmembranes of connective tissue in fat layers always has higher intensity values than poorlyreflective tissues.3. A main peak is always found in the intensity histogram and it is always much higher thanthe secondary peak (if any).Rosin’s thresholding is simple but assumes “here is one dominant population in the imagethat produces one main peak located at the lower end of the histogram relative to the secondarypopulation.”[82]. To locate the threshold in Figure 2.10, a straight line AB is drawn from themain peak A to point B which is the first empty bin of the histogram following the last filled bin.The threshold is located by finding the maximum perpendicular distance from straight line ABas shown in Figure 2.10(a). In addition, we also consider bimodal cases (Figure 2.10(b)) wherehistogram intensity values may be larger than line AB. Since the threshold is most likely locatedat a concavity, histogram values that are greater than AB are not considered. We use geometryto solve the threshold and the calculation is shown in Appendix C. Since the histogram appearsnoisy, it is smoothed by a 5-point Gaussian filter before finding the threshold by Rosin’s method.Chapter 2. Method in Developing Image Processing and Boundary Detection 42(a)(b)Figure 2.10: Detection of threshold using Rosin’s thresholding method on (a) unimodal histogramand (b) bimodal histogram.2.2.2 Thresholding resultsFigures 2.11 to 2.13 show the thresholding results of fc, σ2s and IBS. The values of parametricimages are normalized between 0 to 1. Higher gray intensity values of the histogram indicatebrighter pixels in the parametric images.Rosin’s thresholding is not accurate in detecting the fat boundary in the histogram of fc becauseof the poor contrast in the value of fc. Although the boundary can be detected in fc when thehistogram of fc is bimodal (Figure 2.11), the thresholding method fails to detect the fat boundaryin Figure 2.13. In the histogram of fc in Figure 2.13, the main peak is wide and not obvious. Asobserved in the B-mode image, the value of fc appears to be disrupted by the connective tissuewithin the fat tissue and the resulting parametric image of fc has a poor contrast. Based on ourChapter 2. Method in Developing Image Processing and Boundary Detection 43observations on the fat tissues with different structures, the parametric image of σ2s has a bettercontrast than that of fc and is more suitable for the thresholding technique.The histogram of σ2s shows that the gray intensity pixel values of the fat tissue change morerapidly than that of the non-fat tissue. In the histogram of σ2s , the left side of the vertical thresholdline are pixels that belong to the non-fat tissue and the right side of the vertical threshold line arepixels that belong to the fat tissue. The pixels in the fat region (which are brighter in the grayintensity) have a wider range of gray intensity values than that of the non-fat region (which aredarker in the gray intensity). The pixels from the non-fat region make up a main peak in thehistogram and the pixels from the fat region make up a tail in the histogram. The main peakcorresponds to a slow change in pixel values and the long tail corresponds to a fast change in pixelvalues. The above condition satisfies the assumptions of the Rosin’s thresholding method and theRosin’s thresholding method is able to detect the change from the fat tissue to the non-fat tissuefrom σ2s .We suggest that a relatively rapid change in the pixel values of σ2s at the subcutaneous fatlayer than its other tissue layers is observed because the fat tissue constitutes two very differentcomponents: the fat cells and connective tissue. We discussed the structure of the subcutaneousfat in Chapter 1 and showed that the fat cells in the subcutaneous fat layer are held by connectivetissue. The large-scale variation in speed of sound between fat and connective tissue ( ∼1480ms−1vs ∼1631ms−1) imposes more fluctuations in the power spectrum and can distort the ultrasoundwavefront. As suggested by Hinkelman at el. [66], the subcutaneous fat has greater energy level andwaveform distortion than muscle when they investigated the effect of abdominal wall morphologyon ultrasonic pulse distortion. Therefore, we imply that the high variation in the tissue structureof subcutaneous fat causes more fluctuations in the received spectrum, and this results in a greaterchange in the spectrum variance.Moreover, the thresholding method is able to locate strongly reflective interfaces from the IBS(right most column of figures 2.11 to 2.13). Not only is the fascia located in the binary map ofIBS, structures with strong echoes are also located in the middle of the binary map of triceps andthigh (Figure 2.12 and 2.13). The effects of thresholding will be discussed further in Chapter 4 inthe concept of the full system.Chapter 2. Method in Developing Image Processing and Boundary Detection 44(a)(b)Figure 2.11: Results of thresholding on spectrum properties captured at the human suprailiac site.(a) B-mode image. (b) Spectrum properties’ images(1st row), their histograms (2nd row) and binarymaps (3rd row). The vertical line in the histogram indicates the calculated threshold.Chapter 2. Method in Developing Image Processing and Boundary Detection 45(b)(a)Figure 2.12: Results of thresholding on spectrum properties captured at the human triceps site. (a)B-mode image. (b) Spectrum properties’ images(1st row), their histograms (2nd row) and binarymaps (3rd row). The vertical line in the histogram indicates the calculated threshold.Chapter 2. Method in Developing Image Processing and Boundary Detection 46(a)(b)Figure 2.13: Results of thresholding on spectrum properties captured at the human thigh site.(a) B-mode image. (b) 2nd to 4th row: a spectrum property’s image, histogram and binary maprespectively. The vertical line in the histogram indicates the calculated threshold.Chapter 2. Method in Developing Image Processing and Boundary Detection 472.3 Fat Boundary DetectionBoundary detection is applied to the binary image obtained from thresholding. Since the binaryimage consists of holes and discontinuous edges due to inhomogeneity of the subcutaneous fat andfibrous connective tissues, a line fitting method called Random Sample Consensus (RANSAC) [84]is proposed to reject outliers and link the fat boundary candidates. First, boundary candidatesare obtained from the binary image by differentiation. Then, RANSAC is used to fit the boundarycandidates by assuming the fat boundary is a straight line. The assumption is based on the factthat while the fat boundary is naturally a curve, it approximates a straight line in the narrow fieldof a sagittal ultrasound image.2.3.1 Extraction of Boundary CandidatesThe binary map (BM) consists of only ones and zeros; therefore, boundary candidates can belocated at the transition from zero to one, or from one to zero. In both vertical and horizontaldirections, a transition happens when there is a difference in the binary values between the currentand the next pixel and differentiation can be used to extract boundary candidates.In the binary map of σ2s (BMσ2s ), the possible fat region consists of ones, and the non-fat regionconsists of zeros; therefore, we assume fat boundary candidates happen when a transition from oneto zero in the vertical direction and we do not consider edge pixels with zero to one transition inthe vertical direction. Equation 2.8 is used to compute the likelihood of a pixel being a boundarycandidate in a BMσ2s and is separated into three cases. Let BMσ2s (i, j) be a pixel in the binarymap at the horizontal and vertical coordinates (i, j) and (x, y)boundaryCandidate be the coordinatesof any boundary candidate. A boundary candidate exists at the coordinates (x, y)boundaryCandidateunder the following conditions:(x, y)boundaryCandidate =(i, j − 0.5) if BMσ2s (i, j)−BMσ2s (i, j − 1) = −1(i+ 0.5, j) if BMσ2s (i+ 1, j)−BMσ2s (i, j) = 1(i+ 0.5, j) if BMσ2s (i+ 1, j)−BMσ2s (i, j) = −1(2.8)Chapter 2. Method in Developing Image Processing and Boundary Detection 48(a) (b)Figure 2.14: An example illustrates the extraction of boundary candidates in (a) a binary map ofσ2s using Equation 2.8 and (b) a binary map of IBS using Equation 2.9. A white pixel representsthe value of one and a black pixel represents a pixel of zero. Blue solid crosses denote the bound-ary candidates obtained in the vertical direction and green dotted crosses denote the boundarycandidates obtained in the horizontal direction.Referring to the first case of Equation 2.8, a transition of a pixel valued one to a pixel valuedzero in the vertical direction denotes an existence of a boundary candidate in between the twovertically adjacent pixels. In the second and third cases, a transition of a pixel valued one toa pixel valued zero or a transition of a pixel valued zero to a pixel valued one in the horizontaldirection denotes an existence of a boundary candidate in between the two horizontally adjacentpixels. Since a boundary candidate is defined at the transition between two pixels, its coordinates(x,y) is recorded at the mid point between pixels. Figure 2.14(a) illustrates the usage of Equation2.8.In the binary maps of IBS (BMIBS), strongly reflective interfaces appear as lines with certainthicknesses. Since the thickness of fascia is human dependent and the usual practice to find the fatthickness is from the skin to the surface of fascia [13], we assume fat boundary candidates are foundat transition from zero to ones in the vertical direction and do not consider edge pixels with one tozero transition in the vertical direction. Equation 2.9 is used to compute the likelihood of a pixelbeing a boundary candidate in a BMIBS and is separated into three cases. Let BMIBS(i, j) be apixel in the binary map at the horizontal and vertical coordinates (i, j) and (x, y)boundaryCandidateChapter 2. Method in Developing Image Processing and Boundary Detection 49be the coordinates of any boundary candidate. A boundary candidate exists at the coordinates(x, y)boundaryCandidate under the following conditions:(x, y)boundaryCandidate =(i, j − 0.5) if BMIBS(i, j)−BMIBS(i, j − 1) = 1(i+ 0.5, j) if BMIBS(i+ 1, j)−BMIBS(i, j) = 1(i+ 0.5, j) if BMIBS(i+ 1, j)−BMIBS(i, j) = −1(2.9)Referring to the first case of Equation 2.9, a transition of a pixel valued zero to a pixel valuedone in the vertical direction denotes an existence of a boundary candidate in between the twovertically adjacent pixels. In the second and third cases, a transition of a pixel valued one toa pixel valued zero or a transition of a pixel valued zero to a pixel valued one in the horizontaldirection denotes an existence of a boundary candidate in between the two horizontally adjacentpixels. Since a boundary candidate is defined at the transition between two pixels, its coordinates(x, y) is recorded at the mid point between pixels. Figure 2.14(b) illustrates the usage of Equation2.9.2.3.2 Fitting Boundary Candidates using Random Sample Consensus(RANSAC)Boundary candidates consist of edge points from not only the real fat boundary, but also from thefibrous membranes of connective tissues and noise. A line fitting algorithm is needed to clusterpoints that lie on the same structure. The Hough transform is the most general line fitting al-gorithm, but it suffers from quantization errors and difficulties with noise [85]. Considering ourboundary candidates, it is noticed that the number of unwanted candidates highly depends on thebody location, and the homogeneity and thickness of subcutaneous fat (e.g. more unwanted candi-dates are found in fat with thicker fibrous connective tissue and when the fat thickness is small.) Itis difficult to pick a suitable grid size of the accumulator array for all different cases. Also, Forsythand Ponce [85] have demonstrated this algorithm is very sensitive to noise. To fit a boundary inthe presence of many outliers, the RANSAC algorithm is proposed due to its robustness to outliers.Chapter 2. Method in Developing Image Processing and Boundary Detection 502.3.2.1 Total Least Squares FittingSince the measurement error depends on the coordinate frame, total least squares fitting is usedinstead of using the classic least squares fitting. Although the least squares fitting is simple, itmeasures only the vertical distance error. Alternatively, the total least square fitting measures theperpendicular error which is more robust to pixel errors. The problem of total least square fittingis to minimize the sum of the perpendicular distances between points and lines, i.e.∑(axi + byi + c)2and its minimization problem can be solved with the Lagrange multiplier and the final solution[85] is:c = −ax− by (2.10)A ab = µ ab (2.11)where A = x2 − xx xy − xyxy − xy y2 − yyEquation 2.11 is a 2D eigenvalue problem. Matlab function eig is used to find the two eigenvaluesd and their eigenvectors −→v (v1, v2) of matrix A and gives ab = d v1v2Finally, a, b and c can be solved by choosing the eigenvalue d which gives the smallest∑(axi + byi + c)2.2.3.2.2 Theory and ImplementationRANSAC randomly picks and fits n data points, and checks how many data points can fit to amodel. The process is iterative and continues until a high probability of finding the correct modelis attained [84]. Given our model (i.e. the fat boundary) is a line structure, our problem is to fitthe edge candidates to a line whose equation is ax+ by + c = 0. The algorithm is outlined in theChapter 2. Method in Developing Image Processing and Boundary Detection 51following:Loop until k iterations have occurred1. Select n boundary candidates at random.2. Fit the set of n data points and find a line model using total least squares fitting.3. If determinant of matrix A (from Equation 2.11) is not zero:To find boundary candidates that are close to the line model:For each boundary candidateTest the perpendicular distance (dist) to the line model,If dist ≤ t,Keep boundary candidateEnd IfEnd ForEnd If4. If (number of candidates) > (current maximum number of good candidates)Save the current sets of boundary candidates.Update k by using equation (2.12).EndIfEnd LoopIn the algorithm, three parameters are needed to be determined and they are:· n is the smallest number of points required to fit the model.· t is the threshold (in pixel) required to determine if the data fit well.· k is the number of iterations required for the algorithm.Parameter n is dependent on the fitting model. In our case, n is set to 2 because only twopoints are required to fit a straight line. Occasionally, two randomly picked candidates may fail tofit the line model (e.g. if two candidates are too close to each other.) To ensure the fitting model isvalid, the determinant of matrix A is calculated and checked if it is zero. If the determinant valueis zero, the two candidates cannot generate a valid line model.Parameter t determines whether a boundary candidate is sufficiently close to the fitting modeland represents the maximum perpendicular distance from a good candidate to the fitting line. Thisparameter can be decided by varying the pixel values (in our case, we tried between 0.5 and 2.5pixels) in our data sets and visually determining the best fit. By trial and error in our experiments,t is set to 1 pixel.Chapter 2. Method in Developing Image Processing and Boundary Detection 52Parameter k can be found by considering the probability of k consecutive failures and theprobability of good fit of a random data (Pfit) (Pfail) [85]:Pfail = (1− Pfitn)kk =log(Pfail)log(1− Pfitn)(2.12)We assume the probability of a successful fitting is 99% and therefore Pfail is set to 0.01. k isupdated whenever a better set of boundary candidates is found.Figures 2.15, 2.17 and 2.19 show the results of boundary detection on σ2s and Figures 2.16, 2.18and 2.20 show the results of boundary detection on IBS. In the figures, sub-figure(a) shows theboundary candidates extracted by the differentiation method as described in section 2.3.1 and sub-figure(b) shows the boundary detected by RANSAC. The RANSAC algorithm successfully omitsthe false boundary candidates caused by holes within the fat tissue area (i.e. binary value equals1) and pixels from noise.(a) (b)Figure 2.15: Results of extraction and detection of boundary candidates from the binary mapBMσ2s obtained from a human suprailiac site: (a) potential boundary candidates (red crosses) and(b) fat boundary candidates found by RANSAC (red crosses).Chapter 2. Method in Developing Image Processing and Boundary Detection 53(a) (b)Figure 2.16: Results of extraction and detection of boundary candidates from the binary mapBMIBS obtained from a human suprailiac site: (a) potential boundary candidates (red crosses)and (b) fat boundary candidates found by RANSAC (red crosses).(a) (b)Figure 2.17: Results of extraction and detection of boundary candidates from the binary mapBMσ2s obtained from a human triceps: (a) potential boundary candidates (red crosses) and (b) fatboundary candidates found by RANSAC (red crosses).Chapter 2. Method in Developing Image Processing and Boundary Detection 54(a) (b)Figure 2.18: Results of extraction and detection of boundary candidates from the binary mapBMIBS obtained from a human triceps: (a) potential boundary candidates (red crosses) and (b)fat boundary candidates found by RANSAC (red crosses).(a) (b)Figure 2.19: Results of extraction and detection of boundary candidates from the binary mapBMσ2s obtained from a human thigh: (a) potential boundary candidates (red crosses) and (b) fatboundary candidates found by RANSAC (red crosses).Chapter 2. Method in Developing Image Processing and Boundary Detection 55(a) (b)Figure 2.20: Results of extraction and detection of boundary candidates from the binary mapBMIBS obtained from a human thigh: (a) potential boundary candidates (red crosses) and (b) fatboundary candidates found by RANSAC (red crosses).2.3.3 Calculation of Spectral Content using Multiple FocusesThere is only one focus available in the RF data capture mode of our ultrasound machine. How-ever, fat thicknesses can be different among people and the focus at a fixed position may not beappropriate for all thicknesses. For the above reason, the use of multiple focal points is considered.RF data is separately captured for different focuses and converted to spectrum properties by spa-tial compounding. Two approaches are investigated to combine spectrum properties obtained bymultiple focuses: stitching and averaging.2.3.3.1 Stitching Focused Spectrum Properties (MF1)The approach of stitching combines a spectrum property obtained from several focuses as shownin Figure 2.21(a). This is an idea of combining the focused spectrum property values from pairsof focuses. The region of focus Fx is overlapped with the next focus Fx+1. To guarantee a smoothtransition, the two regions are combined by a ramp-like weight function as shown in Figure 2.21(b).In the overlapping region, the sum of the weights from region Fx and Fx+1 at a certain depth is100%, and the weight of region Fx decreases when the depth increases and the weight of regionFx+1 increases when the depth increases. Focuses Fx at 10mm, 20mm, 30mm and 40mm are usedin our experiments.Chapter 2. Method in Developing Image Processing and Boundary Detection 56FocusF1Focus F2Focus F3Focus F4Region F1Region F2Region F3RegionF4Lateral DirectionAxialDirection(a)(b)Figure 2.21: Stitching of a spectrum property map obtained from multiple focuses: (a) stitching ofspectrum properties values (b) the weight function that combines two overlapping regions.2.3.3.2 Averaging Spectrum Properties from Multiple Focuses (MF2)This approach simply averages all the whole parametric images obtained from multiple focuses.The idea is to smooth parametric images and reduce the effect of a particular focus.2.3.3.3 ResultsFigure 2.22 shows the resulting compound parametric images of σ2s and IBS obtained at differentsingle focuses and the result of stitching (MF1) and averaging (MF2) multiple focuses. It is shownthat the results of using the averaging method (MF2) gives a smoother σ2s and IBS and leads toa binary map with fewer holes.Chapter 2. Method in Developing Image Processing and Boundary Detection 57FSF SFFSFFFSF MF1 MF2(a)SF SFFFFFSF SF MF1 MF2(b)Figure 2.22: The comparisons between (a)σ2s and (b)IBS obtained from: (1st-4th column) singlefocuses (SF) where F indicates the focus position, (5th column) stitching spectrum properties frommultiple focuses (MF1) and (6th column) averaging spectrum properties from multiple focuses(MF2).2.4 SummaryWe investigated the characteristics of human subcutaneous fat in terms of the properties of powerspectrum and encoded the values of spectrum properties into gray-scale parametric images. Then,we presented the method of image processing and fat boundary detection on these parametric image.We found that the spectrum properties σ2s and IBS could be used to characterize subcutaneousfat. The parametric image of σ2s represents a coarse area of the subcutaneous fat tissue and theparametric image of IBS represents potential locations of the fascia.The characteristics of fat observed in the parametric images of σ2s and IBS are more distinctChapter 2. Method in Developing Image Processing and Boundary Detection 58after the application of spatial compounding that reduces spectrum noise due to speckle. Theparametric image of σ2s shows the area of subcutaneous fat and the parametric image of IBSshows the potential area of the fascia and other strong tissue reflectors. The histogram of theparametric image of σ2s shows that fat pixels change more rapidly than non-fat pixel and a long tailin the histogram results. We suggested that the relative rapid change in the gray-intensity valuesof fat pixels is due to the high variation in the tissue structure of subcutaneous fat. This causeshigh fluctuations in the received spectrum, and this results a more rapid change in the spectrumvariance.Histograms of σ2s can be bimodal or unimodal; the Rosin’s unimodal thresholding method isused to detect the global threshold which separated the fat and non-fat pixels in the parametricimages of σ2s . The same thresholding method was applied to the parametric images of IBS. Then,boundary candidates were extracted from the binary map and the fat boundary was detected byRANSAC. Total least squares was used as a fitting algorithm and we were able to use a the samefitting threshold (i.e. 1 pixel) to detect the fat boundary at the human suprailiac, triceps and thighsites. Finally, we proposed two methods: the stitching and averaging spectrum properties obtainedfrom multiple focuses to overcome the drawbacks of a single focus.59Chapter 3Experimental MethodologyThis chapter presents an overall framework of our human subcutaneous fat detection method usingthe spectrum properties of RF data, which we discussed in Chapter 2. To test our algorithm,our subcutaneous fat detection framework was applied to human participants at the suprailiac,triceps and thigh sites. These sites were chosen because they have a good range of fat thicknessand they have been popularly used for the assessment of body fat in both females and males[13]. Different properties of fat are also seen in these regions: fat at the suprailiac site is morehomogeneous, whereas fat in the thigh and triceps is less homogeneous and varies in the densityof fibrous connective tissue. Furthermore, these body sites have large regions of flat areas thatallow the linear transducer to make a complete contact with the skin surface without compression.The ultrasound measurements will also be compared to the skinfold caliper measurements. Theprocedures for using ultrasound and skinfold caliper to measure subcutaneous fat are described atthe end of this chapter.3.1 Overview of the Human Subcutaneous Fat DetectionFrameworkThe overview of the human subcutaneous fat detection framework is shown in Figure 3.1. Thedetection process is divided into four main steps: data capture, calculation of spectrum parameters,preprocessing of spectrum properties maps and segmentation.Chapter 3. Experimental Methodology 60Figure 3.1: The framework of human subcutaneous fat detection.Chapter 3. Experimental Methodology 613.1.1 Data captureRF data was captured by using an Ultrasonix ES500 (Ultrasonix Medical Corporation, Burnaby,BC) with a L9-4 linear transducer. The L9-4 linear transducer operates in a frequency rangebetween 4 and 9MHz. In our experiment, the central transducer frequency was set to 6.6MHz. Theresearch package of the ES500 allowed us to build software for direct access to the RF data andcontrol of ultrasound parameters like transducer frequency, field of view and time gain compensation(TGC).RF data capture software was written in Visual C++ and Microsoft Foundation Classes (MFC).A frame of RF data was captured for each steering angle and each focus. The software also allowedusers to control the number and step size of compounding angle and number of focuses. Changingthe focus position was faster than the steering angle, so the fastest sequence of data collection isshown in Figure 3.2. TGC was set to the “Muscleskeleton”preset on the ES500.The data scanning depth was 50mm and the scanning width was 38.2mm. Each frame ofRF data consisted of 127 RF scan lines and each scan line consisted of 2560 data points. 11compounding angles at a 2◦ step size were used for the human experiment. Data were captured ateach focus every 10mm, from 10mm to 40mm.With the beam angle steering function, the current refresh rate in capturing a frame of RF datais 2Hz with the size of RF data stated above. Frames of RF data were saved to files for later offlineprocessing in Matlab.Chapter 3. Experimental Methodology 62If current focalposition > final positionAdvance focalposition by oneincrement- Reset focus to homeposition.- Increase angle by onestep size.If currentangle > final angleEndYesYesNoNoCapture RFdataFigure 3.2: The sequence of capturing RF data.3.1.2 Calculation of Spectrum PropertiesThe second step is to calculate spectrum properties from RF data captured at each steering angle(θ) and focus position (F ). From the experiment mentioned in Section 2.1.2.2.2, we found thatσ2s and IBS were the most indicative factors in determining the fat boundary. Therefore, weonly considered the calculation of σ2s and IBS in our framework. The local power spectrum andtheir spectrum properties are computed as in Section 2.1.1. As a result, the parametric images ofChapter 3. Experimental Methodology 63spectrum properties σ2s(x, y, θ, F ) and IBS(x, y, θ, F ) are computed at each θ and F .3.1.3 Pre-processing of Spectrum Properties MapThe next step is the pre-processing which includes spatial compounding and combining data frommultiple-focused parametric images. The values of σ2s(x, y, θ, F ) and IBS(x, y, θ, F ) at each focusposition are first spatially compounded to the parametric images σ2s(x, y, F ) and IBS(x, y, F ).Spatial compounding converts the parametric images from their data coordinates to the spatialcoordinates and the method was described in detail in Section 2.1.2. After that, the parametricimages captured at different focal positions are combined with the method of Section 2.3.3. Theresultant parametric images of σˆ2s(x, y) and ˆIBS(x, y) are normalized. The normalization rescalesthe pixel intensity values of spectrum properties maps between 0 to 1, and increases the contrastbetween layers. The normalized σˆ2s(x, y) and ˆIBS(x, y) are then smoothed by a 3x3 Gaussian filterto further remove noise. The compounded, normalized and smoothed parametric images of σ2s(x, y)and IBS(x, y) will be used for segmentation.3.1.4 SegmentationLastly, the fat boundary was delineated from the parametric images by Rosin’s thresholding andRANSAC. An intensity histogram of 128 bins was computed from each parametric image. Rosin’sthresholding, which is described in Section 2.2, was used to find the threshold and obtain a binarymap representing the subcutaneous fat layer. The threshold indicates the change from subcutaneousfat to the muscle layer.Potential boundary candidates were obtained from the binary maps (BMσ2s (x, y) andBMIBS(x, y))using the differentiation method defined in Section 2.3.1. RANSAC is then applied to find the fatedges. As mentioned before in Section 2.3.2.2, the parameter n (i.e. smallest number of points re-quired to fit the model) was set to 2, and the parameter t (i.e. the threshold required to determineif the data fit well) was set to 1 pixel. The detected edges of the fat boundary may be incomplete;therefore, cubic spline interpolation was used to link broken edges using Matlab toolbox to obtainthe final fat boundary bBMσ2s(x, y) and bBMIBS (x, y). The reason for choosing the cubic splineinterpolation over the linear interpolation is that the resulting boundary is smoother in the formercase. Figure 3.3 illustrates an example of the segmentation.Chapter 3. Experimental Methodology 64Thresholdthe parametricimages by the RosinThresholding method.Extract potentialboundary candidates.Detect the boundary byRansac.Link broken edges by thespline interpolation.A parametricimage ofA parametricimage ofFigure 3.3: An example illustrates the segementation process on the parametric images of σs andIBS.Chapter 3. Experimental Methodology 653.2 Procedures in User StudyNine volunteers - five females and four males aged between 20 to 30 – were recruited for the userstudy1. The study consisted of two parts: skinfold caliper measurements and ultrasound datacapture. The skinfold caliper test was carried out before the ultrasound test to avoid bias onskinfold measurements.Measurements took place at three body sites: the suprailiac, triceps and thigh areas. For theskinfold caliper and ultrasound measurements, two sets of results were both collected on the leftand right sides of the body as shown in Figure 3.4. The direction of the arrows indicates the grasp ofthe skinfold caliper. At the suprailiac site(Figure 3.4(a)), the measurement was taken at above thecrest of the ilium in a diagonal fold of skin. Measurements were taken midway between the shoulderand elbow at the triceps (Figure 3.4(b)), and midway between the inguinal crease and proximalborder of patella figure(3.4(c)). In addition, two sets of ultrasound data, one set from each side ofthe body, were taken at each site at random locations. As a result, there were 36 sets of ultrasounddata for evaluating the results of segmentation and 18 sets of skinfold and ultrasound data forevaluating the results between skinfold and automatic ultrasound measurements. The proceduresfor the skinfold and ultrasound measurements are presented in the next section. Since our methodis compared to manual segmentation, the procedure of manually locating the fat boundary is alsopresented.3.2.1 Measurement of Skinfold Fat ThicknessPrior to the user study, the investigator2 practiced the skinfold caliper technique until consistentmeasurements were obtained. A Lange caliper (Figure 3.5) was used throughout the experiment.The maximum thickness that can be measured by this caliper is 60mm. The reading was recordedto the nearest 0.5mm.The investigator followed the skinfold measurement technique prescribed by the Canadian So-ciety for Exercise Physiology [86] and was instructed by an experienced skinfold caliper operator3.During the procedures, the participants were asked to stand and relax. A cross, which indicated the1This study was reviewed and approved by the Behavioral Research Ethics Board, Reference No.: B05-0820. Thecopy of the certificate is attached in the Appendix D.2Jessie Ng, author of this thesis.3Barry Legh, Senior Instructor, Human Kinetics Department, UBC.Chapter 3. Experimental Methodology 66(a) suprailiac (b) triceps(c) thighFigure 3.4: Body sites selected for skinfold caliper and ultrasound measurements. The direction ofthe arrows indicates the grasp of the skinfold caliper.Chapter 3. Experimental Methodology 67Figure 3.5: A Lange skinfold caliper.location of the skinfold grip, was first pen-marked on the skin of the site. The skinfold of fat wasgrasped by the thumb and forefinger with one hand 1cm above the pen-mark. The grasped skinfoldwas shaken with both fingers to avoid including the muscle layer in the measurement. With thecaliper in another hand, the caliper was positioned at the pen-mark and released. The pressure offingers should be maintained during the release of the caliper. A reading was taken after the com-plete release of the caliper and the indicator on the caliper had become stable. Measurements wererepeated three times and the final result was the mean of the three measurements. Furthermore,to ensure consistency in maintenance of the pressure by fingers on the skinfold, all measurementswere made by the same investigator.3.2.2 Collection of Ultrasound DataUltrasound data collection was performed after the collection of skinfold measurement data. Thesettings of the ultrasound machine and the type of the transducer used have been detailed in Section3.1.1. To ensure proper contact of the transducer with the skin, and valid capture of data, theB-mode image was used to aid the capture process to obtain good image quality. If the investigatorwas unable to recognize the fat layer, adjustments were made to the position or orientation of thetransducer. Since the fascia is specular in nature, the brightness of the fascia is affected on theincident angle of the ultrasound. The B-mode image can also help the investigator to adjust theChapter 3. Experimental Methodology 68orientation of the transducer for obtaining maximum reflection from the fascia. At the tricepsand the thigh, the transducer was positioned in the sagittal plane. This is because the transduceris a linear array and it is not convenient for it to be placed on an arched surface. Also, tissuecompression can be avoided.Ultrasound gel was applied to the skin to act as a coupling medium between the skin and thetransducer. For each body site, the center of the ultrasound transducer was positioned at thecross. Ultrasound measurements were taken in the sagittal plane for the triceps and thigh. At thesuprailiac site, the transducer was aligned to match the direction of skinfold measurement. Thethickness of the gel was just enough (∼0.1mm) to obtain a clear B-mode image. Excessive gel wasremoved by sweeping the transducer back and forth. Moreover, the transducer was kept uprightto the skin. The investigator applied just enough pressure on the transducer to allow contactbetween its surface and the skin while avoiding the compression of the subcutaneous fat layer.From our experience, the compressibility of fat can be up to approximately 1/3 of the original fatthickness depending on the total fat thickness. This could be determined from the B-mode imageby observing the change in the thickness of fat when it was alternatively compressed and relaxed.3.2.3 Reference Fat Boundary from Manual SegmentationThe automatic segmentation result was compared to that for manual segmentation. The referencefat boundary, which is defined as the thickness between the skin surface and the fascia surface,was obtained by manual delineation on the B-mode image. For each column of the B-mode image,the investigator clicked on the pixel that could be recognized as the fat boundary. The selectedboundary edges were linked together by cubic spline interpolation to obtain a complete boundary.3.3 Evaluation MethodIn our application, we are interested in finding the fascia, which is the subcutaneous fat boundary,and comparing it to our reference boundary obtained by manual segmentation. To evaluate theboundary error between manual segmentation and auto-detection, we used two parameters: theaverage thickness error (dERR) and the root mean square error (dRMS).Chapter 3. Experimental Methodology 69Assume a boundary B consisting of a set of N boundary points, then B can be represented as:B = {(b1x, b1y), (b2x, b2y), ..., (bnx, bny)...(bNx, bNy)}where bnx is the x-coordinate and bny is the y-coordinate of a boundary candidate. We will definethe coordinates of the reference boundary as R = {(r1x, r1y), (r2x, r2y), ..., (rnx, rny), ...(rNx, rNy)}and the segmented boundary as S = {(s1x, s1y), (s2x, s2y), ..., (snx, sny), ...(sNx, sNy)} for calculatingdERR and dRMS in the following sections.3.3.1 Average Thickness Error MetricsSince the average thickness (d) of S isd =∑Nn=1 bnyN(3.1)where bny is the depth of an edge point bn and N is the total number of edge points. Then, theaverage thickness error (dERR) between the reference boundary and the segmented boundary is thesum of the differences between their y-coordinates at the same column of pixels. We define dERR as:dERR =∑Nn=1 (sny − rny)N(3.2)where sny is the y-coordinate of an edge point in the segmented boundary S and rny is the y-coordinate of an edge point in the reference boundary R. A positive dERR indicates the detectedaverage thickness is overestimated and a negative dERR indicates the detected average thickness isunderestimated. In our evaluation, the reference boundary is the boundary obtained from manualsegmentation.Note that d and dERR are in pixels. To convert their unit to mm, we can multiple them by aconversion factor C. If a scan line of RF data consists of Nrf points, the window size of STFT isw and the image depth is D, C can be calculated as:C =DNrf∗w2(3.3)Nrf = 2560, w = 32 and D = 50mm in our experiment. Therefore, C equals 0.31 mm/pixel.Chapter 3. Experimental Methodology 703.3.2 Root Mean Square Error MetricSince the positive and negative errors can cancel in dERR, we also look into the root mean squareerror (dRMS) that tells the deviation of a boundary measurement. To measure the root meansquare difference between the reference and the segmented boundaries, dRMS was used and it canbe calculated as follows:dRMS =√∑Nn=1 (sny − rny)2N(3.4)where sny is the y-coordinate of an edge point in the segmented boundary S, rny is the y-coordinateof an edge point in the reference boundary R and dERR is the average thickness error. The dRMS isthe average distance of a data point from the reference boundary, measured along a vertical distance.In our evaluation, the reference boundary is the boundary obtained from manual segmentation.3.3.3 Difference against MeanIn addition to the evaluation of results between the manual and automatic ultrasound measure-ments, we also compare the average thickness of the ultrasound measurements with the 12 skinfoldcaliper measurements. The 12 skinfold thicknesses is used because the skinfold caliper measures afold of skin that comprises two layers of subcutaneous fat. Since the ultrasound and skinfold aretwo independent measurements and we do not know the true value of fat thickness, the method ofdifference against mean helps us to investigate the possible relationship between the measurementerror and the true value [87].As shown in Figure 3.6, the mean difference D and the standard deviation of differences(s) areused to characterize the difference against mean between the skinfold and ultrasound measurements.The x-axis is the paired mean between the 12 skinfold and ultrasound measurements, and the y-axisis the paired difference between the 12 skinfold and ultrasound measurements. D is the systematicdifference between methods and s is the standard deviation of differences. D±2s indicates the 95%limits of agreement.3.4 SummaryThe overall subcutaneous fat detection framework was discussed. The subcutaneous fat detectioninvolved four main steps: RF data capture, calculation of σ2s and IBS from the local spectrumChapter 3. Experimental Methodology 71Mean thicknessesof1/2 Skinfold and Ultrasoundmeasurements.Differenceinthicknesses(1/2Skinfold-Ultrasound)Figure 3.6: A figure showing the mean difference D and the standard deviation of differences(s)between two methods. Example data are provided for illustration.of RF data, pre-processing of the parametric images of spectrum properties using spatial com-pounding and segmentation. Two detection results are obtained from σ2s and IBS respectively. Toevaluate our automatic segmentation results, the manually detected boundaries from ultrasoundB-mode images were first used as reference boundaries to compare with the automatically detectedboundaries. The average thickness error (dERR) and the root mean square error dRMS were usedto evaluate the manual and automatic results. The second part of the evaluation was to comparethe ultrasound measurements with the skinfold caliper measurements. We finally presented themethod of difference against mean for the comparison between the ultrasound measurements andthe frac12 skinfold caliper values.72Chapter 4Evaluation of ResultsThis chapter presents the results of subcutaneous fat boundary detection using the proposed imag-ing processing and detection framework. In Chapter 2, we showed that compound parametricimages of σ2s indicate the subcutaneous fat region while the compound parametric images of IBSindicates possible location of fascia and other strongly reflective tissue. Therefore, we first illustrateand discuss the qualitative results of segmentation using σ2s . Then the average thickness and theroot mean square thickness errors between the manual and automatic ultrasound measurements areused to quantitatively evaluate our segmentation results. In the quantitative evaluation, the dif-ference between using σ2s and IBS as our segmentation factors is first assessed. Then, we comparethe segmentation results obtained from a single focus (SF), stitching spectrum properties frommultiple focuses (MF1) and averaging spectrum properties from multiple focuses(MF2). Lastly,the correlations between ultrasound (both manual and automatic methods) and the half skinfoldcaliper measurements are presented.4.1 Qualitative Results: Segmentation Using Spectrum Varianceσ2sFor human subcutaneous fat detection framework is applied to human in vivo, this section presentsthe qualitative result of segmentation with σ2s obtained from a single focus. Figures 4.1(a) to4.1(i) illustrate some results from the fat detection algorithm for subcutaneous fat with differentbody sites, thickness and structure. The cyan boundaries represent the result of the manualsegmentation carried out by the investigator prior to the automatic segmentation, and the redboundaries represent the result of the automatic segmentation. The structure of fat tissue ishuman dependent and the fat tissue found at the suprailiac site is more homogeneous than that atthe triceps and thigh. At the triceps and thigh, the structure is more complicated as more fibrousChapter 4. Evaluation of Results 73membranes are dispersed within the fat tissue. The length, density and thickness of fibrous tissuemembranes vary in these examples; however, our fat detection algorithm is able to identify thelocation of the fat boundaries with varying degrees of accuracy.Chapter 4. Evaluation of Results 74Figure 4.1: Examples (1st row: suprailiac, 2nd row: triceps and 3rd row: thigh) demonstrate thesegmentation results of σ2s at a single focus obtained from different structures and thicknesses ofsubcutaneous fat tissue. The cyan boundary is the manual segmentation and the red boundary isthe automatic segmentation.Chapter 4. Evaluation of Results 75It is harder to detect the fascia boundary when the layer of fat is thin. As shown in the B-modeimages of Figures 4.2(a) and 4.2(b), the subcutaneous fat and the location of fascia are not obvious.The skin is close to the fascia and the subcutaneous fat layer consists of dense fibrous connectivetissues. Since only a small amount of fat tissue is observed between the skin and the fat boundary,it is hard to detect the fat boundary. In the above figures, our algorithm falsely detects the fatboundary at the location near the skin. The binary images of σ2s show that there are too manyboundary candidates and this leads to errors in boundary detection using RANSAC.(a) (b)Figure 4.2: Two examples demonstrate the segmentation results of σ2s on participants with fatthickness ≤5mm. Ultrasound data is obtained using a single focus positioned at 25mm. In sub-figures (a) and (b), the left image is the binary image of σ2s and the right image is the B-modeimage. The cyan boundary is the manual segmentation and the red boundary is the automaticsegmentation.4.2 Evaluation of SegmentationResults between the manual and automatic segmentation are evaluated in this section. First, weshow the mean thickness results of σ2s and IBS obtained from using a single focus positioned at25mm and investigate if IBS is an appropriate parameter for locating the fat boundary at variousbody sites. Secondly, we investigate if there is any improvement in the segmentation result on σ2susing multiple focuses.Chapter 4. Evaluation of Results 764.2.1 Results: Spectrum Variance σ2s vs Integrated Backscattering CoefficientIBSIn this section, we investigate if IBS is an appropriate parameter for locating the fat boundary.The fat boundary, which is fascia, is characterized by strong reflection; however, the heteroge-neous texture of fat appears at different anatomical structure and IBS may not be an accuraterepresentation.4.2.1.1 CorrelationThis section shows the correlation of the average thickness between the manual segmentation andautomatic segmentation. The Pearson’s linear correlation coefficient (r) and the linear regressionequation of the data points are shown. The correlation results are analyzed at each body siteseparately.For the segmentation with σ2s , Figure 4.3 shows that the correlation coefficients r are 0.81, 0.71and 0.82 at the suprailiac, triceps and thigh sites respectively. In the case of segmentation usingIBS, Figure 4.4 shows that the correlation coefficients r are 0.77, 0.14 and -0.20 at the suprailiac,triceps and thigh sites respectively. The results show that there is no correlation between themanual and automatic measurements when using IBS as the spectrum property for segmentationat the triceps and thigh; however, high correlation is found at the suprailiac site.Chapter 4. Evaluation of Results 770 5 10 15 20 25 30051015202530y= 0.77x + 2.74r = 0.81Average Thickness of Manual Segmentation (mm)AverageThicknessofAuto-Segmentation(mm)(a)0 5 10 15 20 25 30 35 400510152025303540y = 1.59x -3.31r = 0.71Average Thickness of Manual Segmentation (mm)AverageThicknessofAuto-Segmentation(mm)(b)0 5 10 15 20 250510152025y = 0.94x + 0.04r = 0.82Average Thickness of Manual Segmentation (mm)AverageThicknessofAuto-Segmentation(mm)(c)Figure 4.3: Correlation between manual and automatic measurements using σ2s at the (a) suprailiac(b) triceps and (c) thigh sites. The red dashed line represents the one-to-one relationship, the blueline is the linear regression line of the 36 samples (blue crosses).Chapter 4. Evaluation of Results 780 5 10 15 20 25 30051015202530y= 0.97x + 1.02r = 0.77AverageThicknessofAuto-Segmentation(mm)Average Thickness of Manual Segmentation (mm)(a)0 5 10 15 20 25 30 35 400510152025303540y = 0.63x + 8.68r = 0.14AverageThicknessofAuto-Segmentation(mm)Average Thickness of Manual Segmentation (mm)(b)0 10 20 30 40 5005101520253035404550y = -0.62x + 26.92r = -0.20Average Thickness of Manual Segmentation (mm)AverageThicknessofAuto-Segmentation(mm)(c)Figure 4.4: Correlation between manual and automatic measurements using IBS at the(a) suprail-iac (b) triceps and (c) thigh sites. The red dashed line represents the one-to-one relationship, theblue line is the linear regression line of the 36 samples (blue crosses).Chapter 4. Evaluation of Results 794.2.1.2 dERRAs discussed in Chapter 3, dERR represents the average thickness error between the boundaries forthe manual and automatic segmentation. For each body site, the paired t-test was conducted todetermine whether the mean of dERR obtained from σ2s is different from that obtained from IBSwith a significance level of 0.05.dERR (mm) Paired t-testBody Site Group Mean ± SD t(df=35) p-value CISuprailiac σ2s −0.32 ± 2.99 −1.18 0.24 (-2.59, 0.68)IBS 0.63 ± 3.95Triceps σ2s 1.29 ± 4.30 −2.23 0.032 (-8.62, -0.40)IBS 5.80 ± 11.16Thigh σ2s −0.48 ± 2.76 −5.42 0.000005 (-16.92, -7.69)IBS 11.82 ± 13.93Table 4.1: Average dERR of 36 samples in 9 subjects (4 samples per participant) at the suprailiac,triceps and thigh sites respectively. The paired t-test is used to compare the mean differencebetween σ2s and IBS. t(df=35): t-value of the paired t-test with a degree of freedom (df) of 35.CI : 95% confidence interval (CI) of the statistical mean difference between σ2s and IBS. If the CIdoes not include 0, there is a significant difference between groups. p-value: if a p-value < 0.05, itindicates there is significant difference between the the dERR of σ2s and IBS.Table 4.1 shows the mean and standard deviation of dERR from σ2s and IBS among 36 samplesat each site. The t-values, p-values and confidence intervals that were calculated by the pairedt-test between σ2s and IBS are presented. If the p-value < 0.05 and the 95% confidence intervaldoes not include 0, there is a significant difference between groups. Significant differences in dERRwere only found at the triceps and thigh (p < 0.05). The comparisons of the dERR differencesbetween σ2s and IBS are summarized as follows:1. At all body site, the values of dERR obtained from σ2s were smaller than that obtained fromIBS.2. Using σ2s as the segmentation parameter, the smallest mean of dERR was noticed at thesuprailiac site(-0.32mm±2.99mm). The thigh had a similar mean of dERR (-0.48mm±2.76mm)with the suprailiac site. The largest mean of dERR was noticed at the triceps (1.29mm±4.30mm).3. Using IBS as the segmentation parameter, the smallest mean of dERR was also noticed at thesuprailiac site (0.63mm±3.95mm). Compared with the result of the suprailiac site, the dERRChapter 4. Evaluation of Results 80was much higher in the triceps (5.80mm±11.16mm) and thigh (11.82mm ±13.93mm). Itindicates that IBS yields higher variance and average value when detecting fat from differentparticipants at the triceps and thigh than the suprailiac site.4. Significant differences in dERR between σ2s and IBS were noticed at the triceps (t(df = 35)= -2.23, p = 0.032) and thigh (t(df = 35) = -5.42, p = 0.000005) but not at the suprailiacsite(t(df = 35) = -1.18, p = 0.024). The boundaries detected from the IBS property have asignificant larger average thickness error than that detected from σ2s at the triceps and thigh,but not at the suprailiac site.4.2.1.3 dRMSAs discussed in Chapter 3, dRMS represents the root mean square thickness error between theboundaries for the manual and automatic segmentation. For each body site, the paired t-test wasconducted to determine whether the mean of dRMS obtained from σ2s is different from that obtainedfrom IBS with a significance level of 0.05.dRMS (mm) Paired t-testBody Site Group Mean ± SD t(df=35) p-value CISuprailiac σ2s 2.00 ± 2.49 −0.03 0.98 (-1.60, 1.55)IBS 2.02 ± 3.49Triceps σ2s 2.38 ± 3.88 −2.32 0.026 (-8.54, -0.57)IBS 6.94 ± 10.50Thigh σ2s 2.10 ± 1.90 −4.78 0.000031 (-15.21,-6.14 )IBS 12.78 ± 13.04Table 4.2: Average dRMS of 36 samples in 9 subjects (4 samples per participant) at the suprailiac,triceps and thigh sites respectively. The paired t-test is used to compare the mean differencebetween σ2s and IBS. t(df=35): t-value of the paired t-test with a degree of freedom (df) of 35.CI : 95% confidence interval (CI) of the statistical mean difference between σ2s and IBS. If the CIdoes not include 0, there is a significant difference between groups. p-value: if a p-value < 0.05, itindicates there is significant difference between the the dRMS of σ2s and IBS.Table 4.2 shows the mean and standard deviation of dERR obtained from σ2s and IBS among 36samples at each body site. If the p-value < 0.05 and the 95% confidence interval does not include0, there is a significant difference between groups. Significant differences in dRMS were only foundat the thigh and the triceps (p < 0.05) . The comparisons of the dRMS difference between σ2s andIBS are summarized as follows:Chapter 4. Evaluation of Results 811. The dRMS values of the triceps (2.38mm±3.88mm) and thigh (2.10mm±1.90mm) that wereobtained from σ2s were smaller than that their values (triceps: (6.94mm±10.50mm) and thigh(12.78mm±13.04mm) obtained from IBS . At the suprailiac site, dRMS obtained from σ2s(2.00mm±2.49mm) was similar than that obtained from IBS (2.02mm±3.49mm).2. Using σ2s as the segmentation parameter, the smallest mean of dRMS was found at the suprail-iac site (2.00mm±2.49mm). The thigh (2.10mm±1.90mm) had a similar mean of dRMS withthat of the suprailiac site. Triceps (2.38mm±3.88mm) had the largest dRMS among all bodysites.3. Using IBS as the segmentation parameter, the smallest mean of dRMS was also found at thesuprailiac site(2.02mm±3.49mm). This value was similar to the error of σ2s (2.00±2.49mm).4. Significant differences in dRMS between σ2s and IBS were noticed at the triceps (t(df = 35) =-2.32, p = 0.026)and thigh (t(df = 35) = -4.78, p = 0.000031) only. The boundaries detectedfrom the IBS property have a significant larger root mean square thickness error than thatdetected from σ2s at the triceps and thigh, but not at the suprailiac site.4.2.2 Results: Multiple-focuses vs Single FocusesTo investigate whether the application of multiple focuses would improve the segmentation resulton σ2s , a one-way ANOVA was applied to dRMS and dERR obtained from the single focus at25mm(SF), stitching multiple focuses(MF1) and averaging multiple focuses(MF2) respectively. Inorder to identify difference between groups, Tukey’s honestly significant difference (HSD) multi-comparison tests were conducted for the pairwise comparisons between SF and MF1, MF1 andMF2, and SF and MF2. Multiple comparisons were performed using Tukey’s method with Matlabfunction multcompare.4.2.2.1 dERRAs discussed in Chapter 3, dERR represents the average thickness error between the boundariesfor the manual and automatic segmentation. For each body site, the one-way ANOVA test wasfirst conducted to determine whether the mean differences of dERR among SF, MF1 and MF2 areChapter 4. Evaluation of Results 82significantly different with a significance level of 0.05. Table 4.3 summarizes the results of theone-way ANOVA tests.dERR (mm) One-Way ANOVABody Site Group Mean ± SD F(2,105) p-valueSuprailiac SF −0.32 ± 2.99 0.08 0.92MF1 −0.63 ± 5.76MF2 −0.30 ± 2.30Triceps SF 1.29 ± 4.30 5.21 0.0069MF1 −1.79 ± 5.58MF2 0.80 ± 2.63Thigh SF −0.48 ± 2.76 0.02 0.82MF1 −0.75 ± 6.79MF2 −0.09 ± 2.23Table 4.3: Average dERR of 36 samples in 9 subjects (4 samples per participant) at the suprailiac,triceps and thigh sites respectively. The one-way ANOVA test is used to compare the meandifference among results obtained from the SF, MF1 and MF2. F(2,105): F-value of the one-wayANOVA test with a between-groups degree of freedom of 2 and a within-group degree of freedomof 105. p-value: if a p-value < 0.05, it indicates there is a significant difference among the groups.Body Site Group A Group B mean difference 95% CI(mm) (mm)Suprailiac SF MF1 0.32 (-1.95, 2.59)SF MF2 −0.01 (-2.28, 2.26)MF1 MF2 −0.33 (-2.60, 1.94)Triceps SF MF1 3.08 ( 0.64, 5.52)†SF MF2 0.49 (-1.94, 2.93)MF1 MF2 −2.58 (-5.02, -0.15)†Thigh SF MF1 0.26 (-2.21, 2.74)SF MF2 −0.39 (-2.87, 2.09)MF1 MF2 −0.65 (-3.13, 1.82)† 95% confidence interval does not include 0; therefore, there is a significant difference betweengroups A and B.Table 4.4: Tukey’s HSD multiple comparisons for the difference in dERR within a group. SF:single focus at 25mm, MF1: stitching multiple focuses and MF2 averaging multiple focuses. meandifference: the estimated statistical mean difference from the Tukey’s HSD test. CI : 95% confidenceinterval of the statistical mean difference between groups A and B.Table 4.3 shows the mean and standard deviation value of dERR from SF, MF1 and MF2 among36 samples at each body site. If the 95% confidence interval does not include 0, there is a significantdifference between group A and B. Significant differences in dERR were only found at the triceps(p < 0.05). Multiple comparisons were then performed between SF and MF1, SF and MF2, andMF1 and MF2 at each body site. Table 4.4 presents the results of multiple comparions in termsChapter 4. Evaluation of Results 83of the statistical mean difference and the 95% confidence interval. If the confidence interval doesnot contain 0, the difference between two groups is significant. If the confidence interval contains0.0, the difference between two groups is insignificant. Comparisons of dERR among SF, MF1 andMF2 are summarized as follows:1. MF1 had the largest mean and standard deviation of dERR at all body sites. Its valuesof dERR were (-0.63±5.76)mm at the suprailiac site, (0.80±2.63)mm at the triceps and (-0.75±6.79)mm at the thigh.2. MF2 had the smallest mean and standard deviation of dERR at all body sites. Its valuesof dERR were (-0.30±2.30)mm at the suprailiac site, (0.80±2.53)mm at the triceps, and (-0.09±2.23)mm at the thigh.3. Significant differences were found between SF and MF1 (CI = (0.64,5.52)), and MF1 andMF2 (CI = (-5.02,-0.15)) at triceps (F(2,105) = 5.21, p = 0.0069) only. In both cases, MF1had a worse dERR than SF and MF2.4. Although the differences between SF and MF2 were insignificant at the suprailiac (CI =(-2.28,2.26)), triceps (CI = (-1.94,2.93)) and thigh (CI = (-5.02,0.15)) sites, there were im-provements on dERR from using MF2 over SF at all body sites. The values of dERR wereimproved from (-0.32±2.99)mm to (-0.30± 2.30)mm at the suprailiac site; (1.29±4.30)mm to(0.80±2.63)mm at the triceps, and (-0.48±2.76)mm to (-0.09±2.23)mm at the thigh.4.2.2.2 dRMSAs discussed in Chapter 3, dRMS represents the root mean square thickness error of between theboundaries between the manual and automatic segmentation. For each body site, the one-wayANOVA test was first conducted to determine whether the mean differences of dRMS among SF,MF1 and MF2 are significantly different with a significance level of 0.05. Table 4.5 summarizes theresults of the one-way ANOVA tests.Table 4.5 shows the mean and standard deviation value of dRMS obtained from SF, MF1 andMF2 among 36 samples at each body site. If the 95% confidence interval does not include 0,there is a significant difference between group A and B. A significant difference in dRMS was onlyChapter 4. Evaluation of Results 84dRMS (mm) One-Way ANOVABody Site Group Mean ± SD F(2,105) p-valueSuprailiac SF 2.00 ± 2.49 1.74 0.18MF1 2.95 ± 5.04MF2 1.42 ± 2.05Triceps SF 2.38 ± 3.88 2.27 0.11MF1 3.71 ± 4.62MF2 1.94 ± 2.07Thigh SF 2.10 ± 1.90 3.68 0.029MF1 3.89 ± 5.68MF2 1.78 ± 1.46Table 4.5: Average dRMS of 36 samples in 9 subjects (4 samples per participant) at the suprailiac,triceps and thigh sites respectively. The one-way ANOVA test is used to compare the meandifference among results obtained from the SF, MF1 and MF2. F(2,105): F-value of the one-wayANOVA test with a between-groups degree of freedom of 2 and a within-group degree of freedomof 105. p-value: if a p-value < 0.05, it indicates there is a significant difference among the groups.found obtained at the thigh (p < 0.05). Multiple comparisons were then performed between SFand MF1, SF and MF2, and MF1 and MF2 at each body site. Table 4.6 presents the results ofmultiple comparions in terms of the statistical mean difference and the 95% confidence interval. Ifthe confidence interval does not contain 0, the difference between two groups is significant. If theconfidence interval contains 0.0, the difference between two groups is insignificant. We summarizethe results when multiple comparing dRMS among the methods of SF, MF1 and MF2:1. Among all methods, MF1 resulted in the largest mean and standard deviation of dRMS at allbody sites. The values of dRMS were (2.95±5.04)mm at the suprailiac sites, (3.71±4.62)mmat the triceps and (3.89±5.68)mm at the thigh.2. Among all methods, MF2 had the smallest average and standard deviation of dRMS at allbody sites. The values of dRMS were (1.42±2.05)mm at the suprailiac site, (1.94±2.07)mmat the triceps and (1.78±1.46)mm at the thigh. Compared with different body sites, theaverage value was lowest at the suprailiac area and the standard deviation was the lowest atthe thigh.3. A significant difference of dRMS between MF1 and MF2 was only noticed at the thigh (CI= (0.12, 4.11)). The mean value of dRMS obtained from MF2 (1.78±1.46)mm is significantlysmaller than that of MF1 (2.10±1.90)mm. Although there was no significant difference ofdRMS between MF1 and MF2 at the triceps (F(2,105) = 2.27, p = 0.11, CI = (-0.29,3.84)),Chapter 4. Evaluation of Results 85Body Site Group A Group B mean difference 95% CI(mm) (mm)Suprailiac SF MF1 −0.92 (-2.88, 1.05)SF MF2 0.62 (-1.35, 2.58)MF1 MF2 1.53 (-0.43, 3.50)Triceps SF MF1 −1.33 (-3.40, 0.73 )SF MF2 0.44 (-1.62, 2.51 )MF1 MF2 1.78 (-0.29, 3.84 )Thigh SF MF1 −1.79 (-3.79, 0.20 )SF MF2 0.32 (-1.68, 2.31 )MF1 MF2 2.11 ( 0.12, 4.11 )†† 95% confidence interval does not include 0; therefore, there is a significant difference betweengroups A and B.Table 4.6: Tukey’s HSD multiple comparisons for the difference in dRMS within a group. SF:single focus at 25mm, MF1: stitching multiple focuses and MF2 averaging multiple focuses. meandifference: the estimated statistical mean difference from the Tukey’s HSD test. CI : 95% confidenceinterval of the statistical mean difference between groups A and B.this showed a trend toward significance.4. There was no significant difference of dRMS between SF and MF2 at all body sites. Nev-ertheless, there are improvements on the mean and standard deviation of dRMS over SF byusing MF2 at all body sites. The values of dRMS were improved from (2.00±2.49)mm to(1.42±2.05)mm at the suprailiac site; (2.38±3.88)mm to (1.94±2.07)mm at the triceps and(2.10±1.90mm) to (1.78±1.46)mm at the thigh.5. No significant difference was noticed between SF and MF1 at all body sites(CI = (-2.88,1.05)at the suprailiac site, CI = (-3.40,0.73) at the triceps and CI = (-3.79, 0.20) at the thigh).However, the value of dRMS obtained from SF is smaller than that of MF1 at all body sites.The values of dRMS were decreased from (2.95±5.04)mm to (2.00±2.49)mm at the suprailiacsite; (3.71±4.62)mm to (2.38±3.88)mm at the triceps and (3.89±5.68) to (2.10±1.90mm) atthe thigh.4.2.3 DiscussionsWe showed that the segmentation with σ2s (obtained from a single focus at 25mm) was a feasibletechnique to detect the location of the fat boundary at the suprailiac, triceps and thigh sites. How-ever, outliers were observed in the linear correlation plots (Figure 4.3) at all such sites. AlthoughChapter 4. Evaluation of Results 86there was high correlation between the manual measurement and automatic measurement usingσ2s , the presence of outliers indicated that σ2s only provided a coarse estimation on the area of fattissue and was affected by the variation in tissue structures of samples.The evaluations on segmentation with σ2s using a single focus at 25mm showed similar errorsat the suprailiac and thigh sites. The dERR and dRMS were (-0.32±2.99)mm and (2.00±2.49)mmat the suprailiac site, and (-0.48±2.76)mm and (2.10±1.90)mm at the thigh. The worst resultwas shown at the triceps where dERR was (1.29±4.30)mm and the dRMS was (2.38±3.88)mm. Webelieve that the reason for the worst results at the triceps was the presence of denser and thickerfibrous of connective tissue there. Earlier, Bellisari et al. [33] also reported the worst technicalerror was found in the triceps site in females.Moreover, the results of segmentation with σ2s are degraded when the fat thickness is too thin.This is because a smaller amount of fat tissue is present between the fascia and the skin while denseconnective tissue is present. Roche [13] also reported that it was harder to define the boundarybetween the subcutaneous fat and muscle in ultrasound B-mode images due to the presence ofsmaller amounts of intermuscular fat tissue. In our experiment, participants had the smallest meanand range of fat thickness at the triceps (Table 4.7), this is another reason for why the largest dERRand dRMS errors are noticed at the triceps.Reference Thickness (mm) Suprailiac Triceps Thighmean±SD 13.34 ± 4.91 7.76 ± 2.56 9.05 ± 4.18minimum 5.45 3.03 3.24maximum 23.82 13.53 22.13Table 4.7: Summary on reference average thickness of subcutaneous fat collected from 9 participantswith 4 samples for each person at each body site. The reference thickness is obtained by manualsegmentation on ultrasound data.A subcutaneous fat boundary is characterized by strong reflection; however, we observed thatit was not reliable to use IBS to locate the fat boundary at the thigh and triceps but it wasfeasible to detect the fat boundary at the suprailiac site. Referring to Figures 4.4(b) and 4.4(c)of the triceps and thigh, we noticed that outliers were mostly located in a deeper area. This isbecause strong echoes can be generated not only by the fibrous connective tissues within the fatlayer, but also by structures like tendons and bones. Our results show that there was no correlationbetween the manual and automatic measurements when using the IBS property at the triceps andChapter 4. Evaluation of Results 87thigh. Further, we discovered that IBS could be used as a segmentation factor at the suprailiacsite because fat is more homogeneous there than at the triceps and thigh. There are no structureswith strong echoes like tendon and bone around this anatomical site, and fewer and thinner fibrousconnective tissue are found there (as shown in Figure 4.5). No significant differences in dERR anddRMS were noticed between segmentation with σ2s and IBS among the 36 samples at the suprailiacsite.(a)(b)(c )Figure 4.5: Show the results of IBS and σ2s at the (a) suprailiac, (b) thigh and (c) triceps sites. IBSis not reliable in locating the fat boundary at the thigh and triceps because of the presence of othersoft tissues with strong reflection. In subfigures (a),(b) and (c), from left to right: binary map fromIBS, segmentation result from IBS, binary map from σ2s and segmentation result from σ2s . Thecyan boundary is the manual segmentation and the red boundary is the automatic segmentation.Chapter 4. Evaluation of Results 88The effect of using multiple focuses to improve results was investigated because we assumedthat thicknesses could be different among people and a focus at a fixed position might not beappropriate for all thicknesses. Our statistical analysis showed that there was no significant benefitin using multiple focuses over a single focus, although we found that averaging σ2s with multiplefocuses smoothed the value of σ2s and tended to report smaller values of dERR (-0.32mm±2.99)mmvs (-0.30±2.30)mm at suprailiac sites, (1.29±4.30)mm vs (0.80±2.63)mm at the triceps and (-0.48±2.76)mm vs (-0.09±2.23)mm at the thigh and dRMS (2.00±2.49)mm vs (1.42±2.05)mm at thesuprailiac,(2.38±3.88)mm vs (1.94±2.07)mm at the triceps and (2.10±1.90)mm vs (1.78±1.46)mmat the thigh, whereas stitching σ2s with multiple focuses degraded the segmentation quality espe-cially in the presence of fibrous tissues.Moveover, Bellisari et al. [33] evaluated the inter-observer technical errors of manual ultrasoundmeasurements and found that the absolute technical errors were 0.15mm at the suprailiac site,0.62mm at the triceps site and 0.13mm at the mid-thigh site. Our absolute mean values of dERRobtained at multiple focuses are close to their technical error. The mean values of dERR were–0.30mm at the suprailiac , 0.80mm at the triceps and -0.09 at the thigh sites. Similarly, both ofour results showed that the worst error was found at the triceps site.From our visual investigation, we noticed that the thresholding result of σ2s captured with asingle focus occasionally underestimated the area of fat tissue. It happened when the surface of thefascia was not well-defined and fibrous tissues appeared near the fascia (as shown in Figures 4.6to 4.8). As a result, a fat layer with significant holes and boundary gaps appeared in the binarymap of σ2s obtained with a single focus (Figure 4.6(a)). However, the gap was annihilated in thebinary map at Figure 4.6(c) after averaging σ2s captured from multiple focuses. Figure 4.7 showsa case where a thin layer of fibrous tissue appeared near the top right of the fascia. Averaging thebinary map of σ2s with multiple focuses reduced the gap near the fat boundary and corrected thelocation of the boundary detected. Figure 4.8 shows another case that benefited from averaging σ2swith multiple focuses. As shown in the B-mode image, an obvious layer of fibrous tissue is foundwithin the layer of fat and the resulting binary map (Figure 4.8(a)) obtained with a single focuswas incorrect. The averaging technique was less sensitive to noise and thick fibrous connectivetissue, and the detection of the fat boundary from the binary map was therefore improved.Chapter 4. Evaluation of Results 89(a)(b)(c)Figure 4.6: An example showing the improvement of using MF2 over SF and MF1. In subfigures (a)SF,(b) MF1 and (c) MF2, from left to right: binary map of σ2s and, segmentation result of σ2s . Thecyan boundary is the manual segmentation and the red boundary is the automatic segmentation.Chapter 4. Evaluation of Results 90(a)(b)(c)Figure 4.7: An example showing the improvement of using MF2 over SF and MF1. In subfigures (a)SF,(b) MF1 and (c) MF2, from left to right: binary map of σ2s and, segmentation result of σ2s . Thecyan boundary is the manual segmentation and the red boundary is the automatic segmentation.Chapter 4. Evaluation of Results 91(a)(b)(c)Figure 4.8: An example showing the improvement of using MF2 over SF and MF1. In subfigures(a) SF,(b) MF1 and (c) MF2, from left to right: binary map of σ2s and segmentation result of σ2s ata triceps. The cyan boundary is the manual segmentation and the red boundary is the automaticsegmentation.Chapter 4. Evaluation of Results 92In addition, we found that the method of stitching spectrum properties with multiple focuses(MF1) resulted in statistically higher values of dRMS (at the thigh) and dERR (at the triceps) thanusing single focus (SF) and averaging multiple focuses (MF2). The idea of stitching the values of aspectrum property at each focus was to improve the detail representation of the spectrum property.However, the presence of extraneous membranes within the fat tissue interfered with the value ofσ2s randomly. Stitching values from multiple focuses exaggerated unwanted details and causederrors in thresholding. On the other hand, averaging the spectrum values obtained from multiplefocuses could smooth out the irregularities and improve the ability to find the global change inthresholding. Although its result did not appear statistically worse at the suprailiac site than theother two methods, we suggest that this may be due to the fact that the structure of subcutaneousfat is different in various body sites. From visual observation, it was noticed that fewer fibroustissues were seen at the suprailiac site than at the triceps and thigh sites. The idea of stitchingthe values of a spectrum property at each focus was to improve the detail representation of thespectrum property. However, the presence of extraneous membranes within the fat tissue interferedwith the value of σ2s randomly. Stitching values from multiple focuses exaggerated unwanted detailsand caused errors in thresholding. On the other hand, averaging the spectrum values obtained frommultiple focuses could smooth out the irregularities and improve the ability of finding the globalchange in thresholding.In the evaluation, we used both the mean thickness error dERR and the root mean squareerror dRMS as the indicator for the segmentation errors because positive and negative y-coordinatevalues can cancel in dERR but not in dRMS . The average thickness errors dERR at the suprailiac,triceps and thigh sites were close to 0mm (< 1mm at all body sites when using the MF2 method).However, the dRMS were all larger than the dERR at all body sites. This implies that averaging theboundary points over the lateral direction reduces the uncertainty of the boundary measurementwhen we are only interested in the average thickness measurement of the fat layer. If we wantto represent the boundary in both axial and lateral direction, the uncertainty of the boundary ishigher than the averaged boundary thickness.Overall, we found that it was more difficult to detect the threshold when the fat layer wasthin or in the presence of thicker, extraneous fibrous membranes. If a fat layer was homogeneouswith less and thinner fibrous membranes, our method would be more robust as σ2s was less noisy;Chapter 4. Evaluation of Results 93however, if fibrous membranes appear in the middle of a homogeneous layer, σ2s was interfered withand changed randomly. The structure of the fat tissue is the main factor of the efficiency of thesegmentation.4.3 Comparison of Auto-detected Fat Thickness with SkinfoldCaliper MeasurementsIn section 4.2, we evaluated the segmentation result in terms of dERR and dRMS and concludedthat the combination of MF2 and σ2s yielded the best result. Therefore, we will use 18 samples innine participants (two sets per participant) of automatic ultrasound measurements from this com-bination for nine human participants to compare with the measurements using skinfold calipers inthis section. Since the investigator was new to the use of the skinfold caliper, an evaluation of herskinfold caliper technique is first presented. Furthermore, the manual and automatic ultrasoundmeasurements are compared with the half of the skinfold thicknesses (12 skinfold thicknesses). The12 skinfold thicknesses is used because the skinfold caliper measures a fold of skin that comprises twolayers of subcutaneous fat. The relationship between the ultrasound and 12 skinfold caliper mea-surements is investigated using linear correlation. After that, the mean difference values betweenthe two techniques are presented.4.3.1 Evaluation on Skinfold Caliper TechniqueA skinfold caliper operator is considered proficient if consistent measurements are made at the samespot. Therefore, we evaluated the skinfold caliper technique of the student investigator by checkingthe discrepancy of measurements that were repeated at a body site. After training and practice, atest was performed on three participants at the suprailiac, triceps and thigh sites. A mark was firstplaced at a body site. In the first trial, three skinfold measurements were taken and averaged toobtain an average thickness of the trial. At an interval of 15 minutes, another three measurementswere recorded and averaged. A total if 5 trials were executed. Water content of subcutaneous fatchanges in time and this may affect the magnitude of the skinfold caliper measurements; therefore,the measurements were taken at the same time of the day at relatively short 15 minute intervals.Table 4.8 presents the justification results in terms of the discrepancy within a trial and theChapter 4. Evaluation of Results 94Site Participant Average Average Discrepancy Average DiscrepancyThickness within a trial from 5 trials(mm) (mm) (mm)Suprailiac 1 6.4 0.2 0.42 23.0 0.4 0.73 24.1 0.5 1.8Triceps 1 6.2 0.2 1.22 18.7 0.5 0.63 11.8 0.2 1.1Thigh 1 6.6 0.3 1.12 29.1 0.8 1.03 6.9 0.2 0.6Overall Median in discrepancy 0.3 1.0Table 4.8: Discrepancies in skin-fold caliper measurement taken at the same spot of a body site.discrepancy of average thickness from five trials. The discrepancy within a trial is the standarddeviation value for the three thickness measurements in the trial. The discrepancy from 5 trialsis the standard deviation value for the averaged thickness from 5 trials. The average discrepancywithin a trial is from 0.2mm to 0.8mm with a median at 0.3mm. Average discrepancy from fivetrials ranged from 0.4mm to 1.8mm with a mean of 1.0mm.The above results are acceptable because the median in discrepancy within a single trial was0.3mm that was within the instrument error of the caliper (i.e. ±0.5mm). Furthermore, the medianin discrepancy from five trials was 1.0mm that was higher than the error of a single trial. This resultwas reasonable because there were more variations between independent trials and the resultingerror still fell at the limit of the caliper resolution (i.e. 1.0mm).4.3.2 Result of CorrelationThe average thickness of the ultrasound boundary is compared to that established by the 12 skinfoldmeasurements. Correlation between 12 skinfold caliper and ultrasound measurements is computedby Pearson’s linear coefficient. 12 skinfold caliper measurements are compared with ultrasoundmeasurements obtained from manual segmentation and automatic detection respectively. Thelinear relationship is presented by both the linear coefficient (r) and the equation of a regressionline. Figures 4.9 to 4.11 shows the linear relationship at the suprailiac, triceps and thigh sitesrespectively. The solid blue line shows the linear relationship between caliper measurement andultrasound measurement, and the dotted blue line represents an one-to-one relationship. The meanChapter 4. Evaluation of Results 95values, linear coefficient and equation of regression line for the three subcutaneous fat measurementsobtained by skinfold caliper and ultrasound techniques are summarized in Tables 4.9 to 4.10.As observed in the case of manual segmentation vs 12 skinfold caliper measurement, high cor-relations are found at the suprailiac (r = 0.93), triceps (r = 0.86) and thigh (r = 0.87) sites.The correlation coefficients r are smaller in the case of automatic detection vs 12 skinfold calipermeasurement and the values are 0.90, 0.72 and 0.89 at the suprailiac, triceps and thigh sites re-spectively. Automatic detection gives a smaller value of r at the suprailiac site and triceps but alarger value at the thigh.Site Caliper (mm) Ultrasound (mm) r Linear RelationshipSuprailiac 12.08 ± 5.17 13.39 ± 5.03 0.93∗ y = 0.90x + 2.47Triceps 5.78 ± 1.92 7.81 ± 2.34 0.86∗ y = 1.05x + 1.74Thigh 10.14 ± 4.56 9.73 ± 4.63 0.87∗ y = 0.88x + 0.83*p < 0.0001Table 4.9: The correlation coefficient r of average thickness between the manual ultrasound seg-mentation vs 12 skinfold caliper measurements for 18 samples in nine participants (two samples perparticipant).Site Caliper (mm) Ultrasound (mm) r Linear RelationshipSuprailiac 12.08 ± 5.17 12.81 ± 3.85 0.90∗ y = 0.78x + 4.00Triceps 5.78 ± 1.92 8.59 ± 2.62 0.72† y = 0.99x + 2.84Thigh 10.14 ± 4.56 9.38 ± 5.12 0.89∗ y = 1.00x - 0.74*p < 0.0001,†p < 0.001Table 4.10: The correlation coefficient r of average thickness between the automatic ultrasoundsegmentation vs 12 skinfold caliper measurements for 18 samples in nine participants (two samplesper participant).Chapter 4. Evaluation of Results 960 5 10 15 20 250510152025y= 0.90x + 2.47r = 0.93UltrasoundAutomaticDetection(mm)1/2 Skinfold Caliper Value (mm)(a)0 5 10 15 20 250510152025y = 0.78x + 4.00r = 0.901/2 Skinfold Caliper Value (mm)UltrasoundAutomaticDetection(mm)(b)Figure 4.9: The relationship of the average thickness between the ultrasound and the skinfoldmeasurements at the suprailiac site: (a) manual ultrasound detection vs 12 skinfold (b) automaticultrasound detection vs 12 skinfold.Chapter 4. Evaluation of Results 970 2 4 6 8 10 12024681012y= 1.05x + 1.74r = 0.86UltrasoundAutomaticDetection(mm)1/2 Skinfold Caliper Value (mm)(a) Manual Ultrasound Detection vs 12skinfold0 2 4 6 8 10 12 1402468101214y = 0.99x + 2.84r = 0.72UltrasoundAutomaticDetection(mm)1/2 Skinfold Caliper Value (mm)(b) 12skinfold vs Automatic DetectionFigure 4.10: The relationship of the average thickness between the ultrasound and the skinfold mea-surements at the triceps : (a) manual ultrasound detection vs 12 skinfold (b) automatic ultrasounddetection vs 12 skinfold.Chapter 4. Evaluation of Results 980 5 10 15 20 250510152025y= 0.88x + 0.83r = 0.87UltrasoundAutomaticDetection(mm)1/2 Skinfold Caliper Value (mm)(a) Manual Ultrasound Detection vs 12skinfold0 5 10 15 20 250510152025y = 1.00x-0.74r = 0.891/2 Skinfold Caliper Value (mm)UltrasoundAutomaticDetection(mm)(b) 12skinfold vs Automatic DetectionFigure 4.11: The relationship of the average thickness between the ultrasound and the skinfoldmeasurements at the thigh: (a) manual ultrasound detection vs 12 skinfold (b) automatic ultrasounddetection vs 12 .4.3.3 Result of Difference Against MeanTo assess the inter-method differences (i.e. skinfold caliper and ultrasound measurements), themean differences between the methods are calculated. Since we do not know the true value offat thickness, difference against mean helps us to investigate the possible relationship betweenthe measurement error and the true value [87] as discussed in Section 3.3.3. For each data set,the difference is computed between the skinfold thickness and ultrasound thickness. The meandifference (D) and standard deviation of differences (s) are computed.For both manual and automatic methods, the largest mean difference is noticed at the triceps.Chapter 4. Evaluation of Results 99The magnitude of mean difference for the automatic method is slightly higher (<1mm) than thatfor the manual method at all sites. The automatic measurements give a higher standard deviationof difference (s) than the manual measurements.Body site Manual Measurement Automatic MeasurementD(mm) s D s(mm)Suprailiac −1.31 1.94 −0.73 2.62Triceps −2.03 1.21 −2.81 1.81Thigh 0.41 2.39 0.76 2.33Table 4.11: Mean difference and standard deviation values between the ultrasound measurementsand the 12 skinfold thicknesses. D is the mean difference and s is the standard deviation.4.3.4 DiscussionsOur results showed that similar correlation values between the manual ultrasound segmentationvs 12 skinfold caliper measurement, and automatic ultrasound segmentation vs12 skinfold calipermeasurements are noticed at both suprailiac and thigh sites. At the triceps, a lower r was obtainedfrom the automatic detection than from manual detection (r = 0.86 vs r = 0.73). This is under-standable because our segmentation algorithm yields a large error in dERR. Several researchers[2, 32, 35, 36, 37] have compared the correlation between the manual ultrasound and skinfold mea-surements at different body sites. Their results are summarized in Table 4.12. Manual ultrasoundand skinfold measurements are highly correlated (r > 0.7) in past studies except for the studycarried out at the suprailiac site by Stevens-Simon et al. [37]. This discrepancy may be due tothe difficulty in obtaining skinfold caliper measurements at the suprailiac sites of pregnant women.Although we cannot directly compare our correlation value r with that of other researchers becauseof differences in sample size, age and gender of participants, our results show a high correlationbetween automatic ultrasound and skinfold caliper measurements. Other researchers using manualultrasound methods have found that there is also a good correlation between the ultrasound andskinfold methods.As we could not establish the true value of fat thickness, we investigated the mean differenceof average thickness between the skinfold caliper and ultrasound measurements. The manual andautomatic ultrasound measurements were compared to skinfold measurements respectively. Im-provement of D by using automatic segmentation occurs at the suprailiac site. However, the mag-Chapter 4. Evaluation of Results 100Correlation Coefficient rReference ResultsBody site Our automatic Volz and Bullen Fanelli and Stevens Ramirezmeasurements Ostrove et al. Kuczmarski -Simon[32] [35] [36] [37] [2]Suprailiac 0.90 0.86 0.85 0.73 0.52 0.84Triceps 0.72 0.79 0.92 0.81 0.89 0.85Thigh 0.89 0.73 − 0.87 0.73 −Table 4.12: A comparison of correlations between ultrasound measurements and skinfold measure-ments at the suprailiac, thigh and triceps sites in this and past studies. The ultrasound measure-ments are obtained by automatic segmentation in this study, and by manual segmentation in theabove past studies.nitudes of mean difference are similar (<1mm) between the manual and automatic segmentationat suprailiac, triceps and thigh sites.It is of interest to note that Ramirez [2] reported differences (D ± s) of (1.2±2.75)mm at thesuprailiac and (4.1±2.85)mm at the triceps sites. With our automatic method, the differenceof (-0.73±2.62)mm at the suprailiac site is similar to Ramirez’s results; however, our differenceof (-2.81±1.81)mm at the triceps which was different from them. This again indicates that thesegmentation results at the triceps is worse than the results at the suprailiac site. There are noreported values at the thigh by Ramirez.We compared the values established through the manual and automatic ultrasound measure-ments to the 12 skinfold caliper values. Although we cannot quantify the true thickness of subcuta-neous fat, the high degree of correlation show that the automatic measurement is at least as goodas the manual method studied by other researchers.4.4 SummaryWe evaluated the automatic ultrasound measurement technique by comparing its results withthose from the manual measurements and 12 skinfold caliper measurements. This showed thatsegmentation with σ2s was a feasible technique to detect the location of the fat boundary at thesuprailiac, triceps and thigh sites. However, the robustness of segmentation with σ2s was affectedwhen the thickness of fat was small and when there were fibrous membranes of connective tissuesnear the fascia. On the other hand, IBS could be used at the suprailiac sites to detect the fatChapter 4. Evaluation of Results 101boundary, but not at the triceps and thigh. Although using multiple focuses to average spectrumproperties reduced the value of dERR and dRMS at all three body sites, there was no significantimprovement in the results . With the presence of fibrous connective tissue, stitching values frommultiple focuses exaggerates unwanted details and causes errors in thresholding. This methoddid not improve the detection and its result was much worse than simply averaging values frommultiple focuses. Since we could not establish the true thickness of fat, we tested the efficiencyof automatic ultrasound measurement by comparing its results to those of manual ultrasound andskinfold measurements. We found that the mean difference at the suprailiac site was similar tothat established in a previous research; however, our value obtained at the triceps was not similar.This was understandable because the automatic segmentation result obtained from the triceps wasthe worst among all sites.102Chapter 5Conclusions and Future Directions5.1 Summary and ConclusionThis thesis explored the use of ultrasound to automatically detect the boundary of subcutaneousfat in vivo. We discovered that the variance of the spectrum (σ2s) and the integrated backscattercoefficient (IBS) carried information related to the properties of subcutaneous fat. We encoded thevalues of σ2s and IBS into separate gray-intensity parametric images, and show that σ2s representsa coarse area of the subcutaneous fat tissue and IBS represents possible locations of the fascia.Then, we presented a framework to detect human subcutaneous fat in vivo by using the informationof σ2s and IBS. A user study of nine participants was carried out to evaluate our segmentationmethod at the suprailiac, triceps and thigh sites.Our subcutaneous fat detection framework consists of four main steps: data capture, calculationof σ2s and IBS from the local spectrum of RF data using STFT and moment analysis, pre-processingof parametric images using spatial compounding and segmentation. Spatial compounding plays avery important role in our framework by reducing noise of the spectrum properties. The non-compounded parametric images of the spectrum properties appear to be noisy because of thespeckle texture of ultrasound. In addition, the non-compounded parametric images of σ2s alsoappear to be erratic in the presence of fibrous connective tissues in the fat tissue. We showedthat spatial compounding reduced speckle noise of the parametric images and differentiated theproperties of subcutaneous fat. We determined the suitable step size and number of steering anglesfrom the experiment with a homogeneous phantom. In the experiment, we noticed that the signal-to-noise ratio increased when the number of angles increased and the step size increased. This isbecause more independent data are available. Although increasing the range of angles can improvethe overall compounding effect, the area covered by the full compounding effect is reduced asdepth increased. In our current approach, spectrum properties obtained from in vivo data wereChapter 5. Conclusions and Future Directions 103interpolated and averaged with 11 compounding angles at 0◦,±2◦,±4◦,±6◦ ±8◦ and ±10◦.After spatial compounding, the visualization of fat using the parametric images of σ2s was im-proved and the specular boundaries shown in the parametric images of IBS were more continuous.The improvement in σ2s was particularly noticed at the triceps and thigh sites where connectivetissues were present in the fat layer and they interfere with the value of σ2s . The histogram of theparametric images of σ2s was characterized by a long tail at higher gray levels. The long tail showsthat fat pixels change more rapidly than non-fat tissues. The fat pixels also have relatively highervalues than the non-fat pixels. We suggested that the relative rapid change in the gray-intensityvalues of fat pixels was due to the high variation in the tissue structure of subcutaneous fat. Thiscaused high fluctuations in the received spectrum, and this resulted in a more rapid change in thespectrum variance. Moreover, the histogram of the parametric images can be bimodal or unimodal.Using Rosin’s thresholding method, we were able to separate the fat and non-fat tissue from σ2sand extracted the possible location of the fascia from IBS. Nevertheless, the IBS also representsother structures with strong echoes such as tendon or bone.Our user study showed that the segmentation with the parametric images of σ2s was a feasibletechnique to detect the location of the fat boundary at the suprailiac, triceps and thigh sites.There are two factors affecting the robustness of segmentation using σ2s . First, it is harder todetect the fascia boundary when the layer of fat is too thin because the connective tissues andthe small amount of fat tissue add uncertainties to the value of σ2s . Second, the robustness of thethresholding algorithm decreases in the presence of connective tissues appearing near the fascia.The thresholding algorithm is more robust in homogeneous fat as σ2s appears to be less erratic. Inour user study, similar levels of dERR and dRMS errors were noticed at the suprailiac and thighsites. The worst result was found at the triceps.On the other hand, IBS can be used at the suprailiac site to detect the fat boundary, but it isnot possible at the triceps and thigh. It is because tendon and bone, which are strong reflectors,can be imaged at the triceps and thigh areas. Our segmentation method could not distinguishthem from the IBS.Our statistical analysis showed no significant improvement in the segmentation results on σ2swhen using multiple focuses to average spectrum properties. However, we showed that it reducesthe mean thickness errors and root mean square errors when compared to the results obtained fromChapter 5. Conclusions and Future Directions 104a single focus and it can improve the thresholding results obtained from a single focus when fibrousconnective tissue is present close to the surface of the fascia. Testing on 36 samples for each bodysite, the mean thickness errors dERR are (-0.30±2.30)mm at the suprailiac site, (0.80±2.63)mmat the triceps and (-0.09±2.23)mm at the thigh, and the root mean square thickness errors dRMSare (1.42±2.05)mm at the suprailiac site, (1.94±2.07)mm at the triceps and (1.78±1.46)mm atthe thigh. We also found that our mean values of dERR were close to the inter-observer technicalerrors of manual ultrasound measurements performed by Bellisari et al. [33]. They found that theabsolute technical errors were 0.15mm at the suprailiac site, 0.62mm at the triceps site and 0.13mmat the mid-thigh site. Similarly, both of our results showed that the worst error was found at thetriceps site.The segmentation results on σ2s obtained by averaging multiple focuses were compared to thoseusing skinfold caliper measurements. As we did not know the absolute true value of fat thickness,comparisons were based on the correlation and the mean difference values. High correlation wasnoticed between the skinfold caliper values and those obtained via the manual and automaticultrasound. When our automatic detection results were compared with the results using skinfoldcaliper measurements, we found that there was a high correlation between the two methods and thecorrelation values were 0.90, 0.72 and 0.89 at the suprailiac, triceps and thigh sites respectively. Thecorrelation appeared lowest at the triceps; however, higher and similar correlations were found at thesuprailiac and thigh sites. Our results showed a high correlation between ultrasound and skinfoldcaliper measurements, which were similar to other researchers’ results. Moreover, the magnitudesof mean difference were similar (< 1mm) when comparing the manual and automatic segmentation.The results indicates that there is a high correlation between our ultrasound measurements andour the skinfold measurements.Although the structure of human subcutaneous fat varies in different body sites and human, ourwork showed that spatial compounded parametric images of ultrasound RF spectrum properties canbe used to segment the subcutaneous fat layer at the suprailiac, triceps and thigh sites of nine humanparticipants. From the histograms of the parametric images of the spectrum variance, we noticedthat the gray-intensity value of fat pixels is higher than that of the non-fat tissues and it changesmore rapidly than those of other tissue layers. We also suggested that this relative rapid change isdue to the fat tissue being composed of two structures with very different acoustic properties: theChapter 5. Conclusions and Future Directions 105fat cells and fibrous connective tissue. Based on our visual observation on the parametric images ofthe spectrum variance, the segmentation algorithm using the thresholding and RANSAC boundarydetection was designed to extract the subcutaneous fat and to calculate the average fat thickness.The main contribution of this work is that an automated technique for determining the humansubcutaneous fat layer using clinical ultrasound has been developed and applied to the humansuprailiac, thigh and triceps in vivo. Our evaluations with the skinfold caliper measurements gavecomparable results with the manual ultrasound measurements previously studied.5.2 Future directionsThe current subcutaneous fat segmentation technique depends in part on the transducer positioningskills of the operator. For instance, the operator has to position the transducer upright to the skinfor maximum reflection from the fascia, and must avoid compression of the fat through observationof the B-mode images. Moreover, the operator must avoid arched surfaces by experimenting withdifferent body locations that allow easy placement of the transducer. A transducer with a smallerfootprint may be a better way to solve the above problems as it covers a smaller skin area. However,the tradeoff is the decrease in field of view of the transducer.The spectrum variance σ2s is more sensitive to the thick, long fibrous connective tissues locatedin particular near the fascia and when the fat layer is too thin. Therefore, further investigationshould be made to reduce the impact of analysis errors by averaging the error effect with a largenumber of analyzed images. Fibrous connective tissues dispersed within the fat tissues appearshorter in length, smaller in volume and less continuous than the fascia which is a continuous sheetof tissue. Viewing shorter and less continuous tissue at different angles or positions will generatedifferent appearances. In our current spatial compounding technique, the direction of beam steeringis parallel to the image plane and the average spectrum properties are only averaged in one plane.However, if we translate the transducer in the direction parallel to the skin (Figure 5.1), we canaverage spectrum properties from adjacent plans and smooth out the effect of fibrous connectivetissues. This idea is similar to the 3D freehand ultrasound. Moreover, this idea can be extendedto volumetric measurement of subcutaneous fat.Further, the experiment is conducted with the ultrasound frequency at 6.6MHz and the thickestChapter 5. Conclusions and Future Directions 106Directionparallel to the skinImage planesFigure 5.1: Moving the transducer parallel to the skin to generate cross sectional images in avolume.sampled fat thickness is smaller than 30mm. Since the frequency of ultrasound affects the pene-tration ability, the efficiency of the thresholding algorithm with growing fat thickness is uncertain.More investigations can be conducted to assess the range of fat thickness that can be measuredusing our algorithm.Ultrasound machines assume that the speed of sound in soft tissues (1540ms−1) is constant;however, the speed of sound is variable in different tissues and fat tissue has a relatively low speedof sound (∼1480ms−1)[52, 53]. The variation in speed casts doubt on the accuracy of spatialthickness measurements and the boundary can be displaced by around 4%(i.e.(1540-1480)/1540).The absolute thickness error increases when the fat thickness increases.Other drawbacks of the system are the slow RF data capture rate. The current frame rate forcapturing RF data is about 2Hz but we need 44 RF frames (11 steering angles with 4 focuses each)for the spectrum calculations; therefore, the total data capture time excluding processing of RFdata is about 22 seconds. To reduce capturing time, fewer steering angles with larger angle stepsizes can be used. The potential problem can be the lower probability of receiving echoes and thelimitation of the depth coverage when the angle step size is too large.Chapter 5. Conclusions and Future Directions 107Our current evaluation method relies on the manual segmentation of B-mode ultrasound im-age and its correlation with skinfold measurements. We do not have a gold standard that has abetter accuracy to evaluate the true value of fat thickness and the effect of the speed displacementdescribed in the previous paragraph. It may be worthwhile to consider using other imaging tech-niques, such as CT and MRI, as additional tools for further comparison. It may be possible thatCT and MRI could provide a better gold standard. Further investigation with a gold standard willhelp to validate and improve the segmentation algorithm.108Bibliography[1] J. Matiegka. The testing of physical efficiency. American Journal of Physical Anthropology,4:223–230, 1921.[2] Maria E. Ramirez. Measurement of subcutaneous adipose tissue using ultrasound images.American Journal of Clinical Nutrition, 89:347–357, 1992.[3] R.J. Kuczmarski, M.T. Fanelli, and G.G. Koch. Ultrasonic assessment of body composition inobese adults: overcoming the limitations of the skinfold caliper. American Journal of ClinicalNutrition, 45:717–724, 1987.[4] C. Nordander, J. Willner, G.A. Hansson, B. Larsson, J. Unge, L.Granquist, and S. Skerfving.Influence of the subcutaneous fat layer, as measured by ultrasound, skinfold calipers and bmi,on the emg amplitude. European Journal of Applied Physiology, 89:514–519, 2003.[5] F. Perin, J.C Pittet, S. Schneert, P. Perrier, F. Tranquart, and P. Beau. Ultrasonic assessmentof variations in thickness of subcutaneous fat during the normal menstrual cycle. EuropeanJournal of Ultrasound, 11:7–14, 2000.[6] DW Newcom, TJ Baas, and JF Lampe. Prediction of intramuscular fat percentage in liveswine using real-time ultrasound. Journal of Animal Science, 80(12):3046–3052, 2002.[7] T. Abe, F. Tanaka, Y. Kawakami, K. Yoshikawa, and T. Fukunaga. Total and segmentalsubcutaneous adipose tissue volume measured by ultrasound. Medicine and Science in Sportsand Excercise, 28:908–912, July 1996.[8] Saito K, Nakaji S, Umeda T, Shimoyama T, Sugawara K, and Yamamoto Y. Developmentof predictive equations for body density of sumo wrestlers using b-mode ultrasound for thedetermination of subcutaneous fat thickness. British Journal of Sports Medicine, 37:144–148,2003.Bibliography 109[9] Kenneth J. Ellis. Human body composition: In vivo methods. Physiological Reviews,80(2):649–680, 2000.[10] Scott B. Going. Densitometry. In T.G. Lohman A.F. Roche, S.T. Heymsfield, editor, HumanBody Composition. Human Kinetics, Champaign, IL, 1st edition, 1996.[11] M.A. McCrory, T.D. Gomex, and E.M. Bernauer. Evaluation of a new air displacementplethysmograph for measuring human body composition. Medicine and Science in Sports andExercise, 27:1686–1691, 1995.[12] L.Z. Coppini, D.L. Waitzberg, and A.C.L. Campos. Limitations and validation of bioelectricalimpedance analysis in morbidly obese patients. Current opinion in Clinical Nutrition andMetabolic care, 8(3):329–332, 2005.[13] A.F. Roche. Anthropometry and ultrasound. In A.F. Roche, S.T. Heymsfield, and T.G.Lohman, editors, Human Body Composition. Human Kinetics, Champaign, IL, 1st edition,1996.[14] AS Jackson and ML Pollock. Generalized equations for predicting body denisty of men. BritishJournal of Nutrition, 3:497–504, 1978.[15] AS Jackson and ML Pollock. Generalized equations for predicting body denisty of women.Medicine and Science in Sports and Excercise, 12(3):175–182, 1980.[16] MS Yuhasz. Physical fitness and sports appraisal laboratory manual. University of WesternOntario:London, 1974.[17] C.Orphanidou, L.McCargar, C.L. Birmingham, J.Mathieson, and E.Goldner. Accuracy ofsubcutaneous fat measurement: Comparison of skinfold calipers, ultrasound, and computedtomography. Journal of the American Dietetic Association, 94(8):855–858, April 1994.[18] Timothy G. Lohman. Dual energy x-ray absorptiometry. In T.G. Lohman A.F. Roche,S.T. Heymsfield, editor, Human Body Composition. Human Kinetics, Champaign, IL, 1stedition, 1996.Bibliography 110[19] P. Tothill, A.Avenell, and D.M. Reid. Precision and accuracy of measurements of whole-bodybone mineral: comparisons between hologic, lunar and norland dual-energy x-ray absorptiome-ters. The British Journal of Radiology, 67(804):120–1217, 1994.[20] J.P. Despres, R.Ross, and S.Lemieux. Imaging techniques applied to the measurement ofhuman body composition. In T.G. Lohman A.F. Roche, S.T. Heymsfield, editor, Human BodyComposition. Human Kinetics, Champaign, IL, 1st edition, 1996.[21] J.C. Seidell, A. Oostelee, MAO. Thijssen, and et al. Assessment of intraabdominal and subcu-taneous adbominal fat: relation between anthropometry and computed tomography. AmericanJournal of Clinical Nutrition, 45:7–13, 1987.[22] K. Tokunage, Y. Matsuzawa, K. Ishikawa, and S. Tarui. A novel technique for the determi-nation of body fat by computed tomography. International Journal of Obesity, 7:437–445,1983.[23] H Kvist. Adipose tissue volume determinations in women by computed tomography: technicalconsiderations. International Journal of Obesity, 10:53–67, 1986.[24] J Kobayashi, N Tadokoro N, Watanbe M, and Shinomiya M. A novel method of measuringintra-abdominal fat volume using helical computed tomography. International Journal ofObesity, 26:398–492, 2002.[25] G. Tornaghi, R. Raiteri, C. Pozzato, A. Rispoli, M. Bramani, M. Cipolate, and A. Craveri.Anthropometric or ultrasonic measurements in assessment of visceral fat? a comparative study.International Journal of Obesity, 18:771–775, 1994.[26] G.C. Dooms, H. Hricak, and G. De Geer A.R. Margulis. MR imaging of fat. Radiology,158:51–54, 1986.[27] M.A. Foster, J.M.S. Hutchison, J.R. Mallar, and M. Fuller. Nuclear magnetic resonance pulsesequence and discrimination of high-and low-fat tissues. Magnetic Resonance Imaging, 2:187–192, 1984.Bibliography 111[28] P.A. Hayes, P.J. Sowood, A. Belyavin, J.B. Cohen, and F.W. Smith. Subcutaneous fat thick-ness measured by magnetic resonance imaging, ultrasound and calipers. Medicine and Sciencein Sports and Excercise, 20:303–309, 1988.[29] Robert Ross. Magnetic resonance imaging provides new insights into the characterizationof adipose and lean tissue distribution. Canadian Journal of Physiology and Pharmacology,74(6):778–785, June 1996.[30] V Positano, A Gastaidelli, AM Sironi, MF Santarelli, M Lombardi, and L Landini. An accurateand robust method for unsupervised assessment of abdominal fat by MRI. Journal of MagneticResonance Imaging, 20(4):684–689, 2004.[31] RAD Booth, BA Goddard, and A Paton. Measurement of fat thickness in man - a comparisonof ultrasound harpenden calipers and electrial conductivity. British Journal of Nutrition,4(20):719–1966, 1966.[32] P.A. Volz and S.M. Ostrove. Evaluation of a portable ultrasonoscope in assessing the bodycomposition of college-age women. Medicine and Science in Sports and Excercise, 16:97–102,1984.[33] A. Bellisari, A.F. Roche, and R.M. Siervogel. Reliability of b-mode ultrasonic measurementsof subcutaneous and intra-abdominal adipose tissue. International Journal of Obesity, 17:475–480, 1993.[34] A. Flygare, L. Valentin, P. Karlsland Akeson, C.E. Flodmark, S.A. Ivarsson, and I. Axelsson.Ultrasound measurements of subcutaneous adipose tissue in infants are reproducible. Journalof Pediatric Gastroenterology and Nutrition, 28:492–494, 1999.[35] B.A. Bullen, F. Quaade, E. Olesen, and S.A Lund. Ultrasonic reflection used for measuringsubcutaneous fat in humans. Human Biology, 37:374–384, 1965.[36] M.T. Fanelli and R.J. Kuczmarski. Accuracy of subcutaneous fat measurement: Comparisonof skinfold calipers, ultrasound, and computed tomography. American Journal of ClinicalNutrition, 39(5):703–709, 1984.Bibliography 112[37] C. Stevens-Simon, P. Thureen, J.Barrett, and E.Stamm. Skinfold caliper and ultrasoundassessments of change in the distribution of subcutaneous fat during adolescent pregnancy.International Journal of Obesity, 25:1340–1345, 2001.[38] S. Kolesnik J. Wang, JC Thornton and PN. Pierson. Anthropometry in body composition -an overview. In vivo body composition studies annals of the New York Academy of Sciences,904:317–326, 2000.[39] Physical status: The use and interpretation of anthropometry. Technical Report 854, Geneva,1995.[40] Canadian population health initiative identifies obesity as research priority. Technical Report 3,Ocober/November 2002.[41] Position of the american dietetic association and the canadian dietetic association: Nutritionfor physical fitness and athletic performance for adults. Journal of the American DieteticAssociation, 93(6):691–696, June 1993.[42] Sinning WE. Body composition in athletes. In A.F. Roche, S.T. Heymsfield, and T.G. Lohman,editors, Human Body Composition. Human Kinetics, Champaign, IL, 1st edition, 1996.[43] KA. Beals KA and MM. Manore MM. Nutritional status of female athletes with subclinicaleating disorders. Journal of the American Dietetic Association, 98:419–425, 1998.[44] A. DeLorenzo, P. Deurenberg, M. Pietrantuono, N. Di Daniele, V. Cervelli, and A. Andreoli.How fat is obese? ACTA Diabetologica, Suppl. 1, 40:S254–S257, October 2003.[45] T. Abe, M. Kondo, Y. Kawakami, and T. Fukunaga. Prediction equations for body compositionof japanese adults by b-mode ultrasound. American Journal of Human Biology, 6:161–170,1994.[46] C.A. Glasbey, I. Abdalla, and G. Simm. Towards automatic interpretation of sheep ultrasoundscans. Animal Science, 62:309–315, 1996.[47] Frederick W. Kremkau. Diagnostic Ultrasound: Principles and Instruments. W.B. SaundersCompany, Philadephia, PA, 6th edition, 2002.Bibliography 113[48] Stephen.W. Flax, Norbert J. Pelc, Gary H. Glover, Frank D. Gutmann, and Maurice McLchlan.Spectral characterization and attenuation measurements in ultrasound. Ultrasonic Imaging,5:95–116, 1983.[49] D. Nicholas, C.R. Hill, and D.K. Nassiri. Evaluation of backscattering coefficients for excisedhuman tissues: results, interpretation and associated measurements. UMB, 7:17–28, 1982.[50] Johan M. Thijssen. Ultrasonic speckle formation, analysis and preciessing applied to tissuecharacterization. Pattern Recognition Letters, 24:659–675, 2003.[51] Douglas A. Christensen. Ultrasonic bioinstrumentation. John Wiley & Sons, New York, 1stedition, 1988.[52] S.A. Goss, R.L. Johnston, and F.Dunn. Comprehensive compliation of empirical ultrasonicproperties of mammalian tissues. Journal of the Acoustical Society of America, 64(2):423–457,1978.[53] S.A. Goss, R.L. Johnston, and F.Dunn. Comprehensive compliation of empirical ultrasonicproperties of mammalian tissues part 2. Journal of the Acoustical Society of America, 68(1):93–108, 1980.[54] G. Treece, R. Prager, and A. Gee. Ultrasound attenuation measurement in the presenceof scatterer variation for reduction of shadowing and enhancement. IEEE Transactions onUltrasonics, Ferroelectrics and Frequency Control, 52:2346–2360, 2005.[55] Frederic L. Lizzi, Ernest J. Feleppa, S. Kaisar Alam, and Cheri X. Deng. Ultrasonic spectrumanalysis for tissue evaluation. Pattern Recognition Letters, 24:637–658, 2003.[56] M. Fink, F. Hotteir, and J.F. Cardosa. Ultrasonic signal processing for in vivo attenuationmeasurement: Short Time Fourier Analysis. Ultrasonic Imaging, 5:117–135, 1983.[57] Y.S. Akgul, C Kambhamettu, and M Stone. Automatic extraction and tracking of the tonguecontours. IEEE Transactions on Medical Imaging, 18(10):1035–1045, 1999.[58] SK Setarehdan and JJ Soraghan. Automatic cardiac LV boundary detection and tracking usinghybrid fuzzy temporal and fuzzy multiscale edge detection. IEEE Transactions on BiomedicalEngineering, 46(11):1364–1378, 1999.Bibliography 114[59] Anant Madabhushi and Dimitris N. Metaxax. Combining low-, high-level and empirical domainknowledge for automated segmentation of ultrasonic breast lesions. IEEE Transactions onMedical Imaging, 22(2):155–169, February 2003.[60] A. Hammoude. Edge detection in ultrasound images based on differential tissue attenuationrates. Ultrasonic Imaging, 21(1):31–42, 1999.[61] D Boukerroui, O Basset O, A Baskurt, and Gimenez G. A multiparametric and multires-olution segmentation algorithm of 3-d ultrasonic data. IEEE Transactions on Ultrasonics,Ferroelectrics and Frequency Control, 1(48):64–77, 2001.[62] Igor Dydenko, Denis Friboulet, Jean-Marie Gorce, Jan D’hooge, Bart Bijnens, and Isabelle E.Magnin. Towards ultrasound cardiac image segmentation based on the radiofequency signal.Medical Image Analysis, 7:353–367, 2003.[63] F. Dvignon, F.-F. Deprez, and O. Basset. A parametric imaging apporach for the segmentationof ultrasound data. Ultrasonics, 43:789–801, 2005.[64] William Bloom and Don W. Fawcett. A textbook of histology. Chapman & Hall, New Yok,12th edition, 1994.[65] Michael H. Ross, Gordon I. Kaye, andWojciech Pawlina. Histology: a text and atlas. LippincottWilliams & Wilkins, Philadelphia, Pa, 4th edition, 2003.[66] L.M. Hinkelman, T.D. Mast, L.A. Metlay, and R.C. Waag. The effect of abdominal wallmorphology on ultrasonic pulse distortion. part i. measurements. Journal of the AcousticalSociety of America, 104(6):3635–3649, December 1998.[67] James F. Greenleaf and Robert C. Bahin. Clinical imaging with transmissive ultrasonic com-puterized tomography. IEEE Transactions on Biomedical Engineering, 28(2):177–185, 1981.[68] L.Landini and R. Sarnelli. Evaluation of attenuation coefficients in normal and pathologicalbreast tissue. Medical & Biological Engienering & Computing, 24(3):243–247, 1986.[69] Uwe Haberkorn, Gu¨nter Layer, Volker Rudat, Ivan Zuna, Adolf Lorenz, and Gerhard vanKaick. Ultrasound image properties influenced by abdominal wall thickness and composition.Journal of Clinical Ultrasound, 21:423–429, 1993.Bibliography 115[70] A. Pomaroli, B. Norer, R. Schellenast, and G. Klima. The sonographic pattern of fatty tissuein the floor of the mouth. Quaderni Di Anatomica Pratica 1, suppl XLIII:49–57, 1998.[71] S. Bridal, P. Fornes, P. Bruneval, and G. Berger. Parametric (intergrated backscatter andattenuation) images constructed using backscattered radio frequency signal (25-56MHz) fromhuman aortae in vitro. Ultrasound Medicine and Biology, 19:215–229, 1997.[72] Gosta H. Granlund and Hans Knutsson. Signal Processing for Computer Vision. KluwerAcademic Pulbilishers, Netherlands, 1st edition, 1995.[73] Nina Eriksoon Bylund, Marcus Ressner, and Hans Knutsson. 3D wiener filtering to reducereverberations in ultrasound image sequences. In Image analysis, Proceedings lecture notes incomputer science, volume 2749, pages 579–586, 2003.[74] Nina Eriksoon Bylund, M. Andersson, and Hans Knutsson. Wide range frequency estimationin ultrasound images. In Proceedings of SSAB Symposium on Image Analysis, march 2001.[75] M. Berson, A. Roncin, and L. Pourcelot. Compound scanning with an electrically steeredbeam. Ultrasonic Imaging, 3:303–308, 1981.[76] S. Jespersen, J. Wilhjelm, and H. Sillesen. Multi-angle compound imaging. Ultrasonic Imaging,20:81–102, 1998.[77] S. Jespersen J. Wilhjelm and H. Sillesen. In vitro spatial compound scanning for improvedvisulization of atherosclerosis. Ultrasonic Imaging, 26(1357–1362), 2000.[78] M. Tanter, J. Bercoff, L. Sandrin, and M. Fink. Ultrafast compound imaging for 2-D motionvector estimation: application to transient elastography. IEEE Transactions on Ultrasonics,Ferroelectrics and Frequency Control, 49:1363–1374, 2002.[79] M. Pernot, M. Tanter, J. Bercoff, K.R. Waters, and M. Fink. Temperature estimaton usingultrasonic spatial compound imaging. IEEE Transactions on Ultrasonics, Ferroelectrics andFrequency Control, 51:606–615, 2004.[80] Haifeng Tu, Tomy Varghese, Ernest L. Madsen, Quan Chen, and James A. Zagzebski. Ultra-sound attenuation imaging using compound acquisition and processing. Ultrasonic Imaging,25:245–261, 2003.Bibliography 116[81] N Otsu. Threshold selection method from gray-level histograms. IEEE Transactions on Sys-tems man and cybernetics, 9(1):62–66, 1979.[82] Paul L. Rosin. Unimodal thresholding. Pattern Recognition, 34:2083–2096, 2001.[83] Paul L. Rosin and J. Hervas. Remote sensing image thresholding methods for determininglandslide activity. International Journal of Remote Sensing, 26(6):1075–1092, 2005.[84] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for modelfitting with applications to image analysis and automated cartography. Communications ofthe ACM, 24(6):381–395, 1981.[85] David A. Forsyth and Jean Ponce. Computer Vision A Modern Approach. Pearson Education,Upper Saddle River, New Jersey, 1st edition, 2003.[86] Canadian Society for Exercise Physiology (CSEP). Canadian Physical Activity, Fitness andLifestyle Approach: CSEP-Health and Fitness Program’s Health-Related Appraisal and Coun-selling Strategy. Canadian Society for Exercise Physiology, 3rd edition, 2003.[87] J. Martin Bland and Douglas G. Altman. Statistical methods for assessing agreement betweentwo methods of clinical measurement. The LANCET, 8:307–310, February 1986.117Appendix ACalculation of the Spectrum CentralFrequency and VarianceFink et al. [56] proved that the the spectrum moments of the short time Fourier transform arecorrelated to the attenuation rate, the spectrum central frequency and variance.For a linear attenuation with frequency and neglecting diffraction effects, Fink et al. [56] showedthat the relationship between the central frequency fc(t) and the spectrum variance σ2s(t) that varywith time t isdfcdt= −βcσ2s(t) (A.1)where dfcdtis the attenuation rate, β is the attenuation coefficient and c is the speed of sound.The spectrum due to attenuation is:ǫ(f, t) = ǫo(t)e−(α(f)ct). (A.2)And if the attenuation is linear with respect to frequency, then Equation A.2 becomesǫ(f, t) = ǫo(t)e−(β|f |ct). (A.3)The nth moment of the spectrum ǫ(f, t) ismn(t) =∫ f1f2ǫo(t)e−(β|f |ct)fndf. (A.4)Appendix A. Calculation of the Spectrum Central Frequency and Variance 118The differentiation of Equation A.4 becomesdmn(t)dt= −βcmn+1(t). (A.5)fc of the spectrum is the spectrum centroid; therefore,fc =m1(t)m0(t). (A.6)dfcdtcan be obtained by differentiating the Equation A.6 by the quotient rule and Equation A.4:dfcdt=ddtm1(t)m0(t)(A.7)= −βc(1m0(t)(dm1(t)dt)−m1(t)m20(t)(dm0(t)dt)) (A.8)= −βc(m2(t)m0(t))− (m1(t)m0(t))2) (A.9)= −βc(m2(t)m0(t))− f2c ) (A.10)= −βcσ2s(t). (A.11)Therefore, the spectrum variance σ2s ism2(t)m0(t)− f2c .119Appendix BNormalization of a SpectrumProperty ImageA parametric image Mij is normalized between 0 to 1 before carrying out the thresholding step.Mij is normalized as follows:Mˆij = g ∗(Mij −min(Mij))(max(Mij)−min(Mij)(B.1)where Mˆij is the normalized Mijg is the maximum value of the normalized parametric image and its value is 1 for normalizing theparametric image between 0 to 1,i is the column index,j is the row index,max is the function to find the maximum value in Mij , andmin is the function to find the minimum value in Mij .120Appendix CSolution to Rosin’s ThresholdingMethod(X , Y )s s(x ,y )z z(x, y) (X , Y )f fHistogramfreqeuncyIntensity ValuexySlope = my = f(x)Figure C.1: Finding the threshold by Rosin’s thresholding method.Let the histogram be a function y = f(x) and the threshold located at the point (x,y) whichhas the maximum perpendicular distance from the straight line in the histogram.Let (Xs, Ys) be the histogram main peak, (Xf , Yf ) the first empty bin of the histogram followingthe last filled bin, m the slope of the straight line joining the (Xs, Ys) and (Xf , Yf ) and (xz, yz)isa point on the straight line that achieves the maximum perpendicular distance to the straight line.To find (xp, yp), the first step is to find equations representing xz and yz.Appendix C. Solution to Rosin’s Thresholding Method 121The slope of the straight line ism =(Ys − Yf )(Xs − Yf ). (C.1)Assume (x,y) is a point on the histogram, its perpendicular distance to the straight line can becalculated by the dot product as follows: x− xzy − yz · Xs − xzYs − yz = 0. (C.2)Rearranging the terms in Equation C.2, we haveax2z + bxz + c = 0 (C.3)wherea = 1 +m2b = −(Xs + x)−m(y − Ys)− 2m2Xsc = xXs +m(XsYs +Xsy − 2XsYs) +m2X2s .xz is solved from the roots of the quadratic Equation C.3 for each point (x,y) that lies betweenmain peak (Xs, Ys) and the last bin (Xf , Yf ).Then, yz can be found asyz = m(xz −Xs) + Ys. (C.4)The perpendicular distance from the straight line to (x,y) is thendistance = (√(x− xz)2 + (y − yz)2). (C.5)The threshold point is the point (x,y) that has the maximum distance in Equation C.5.122Appendix DUBC Research Ethics BoardCertificate of Approval

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0065566/manifest

Comment

Related Items