UNIVERSITY OF CALGARY. In Vivo Assessment of Bone Microarchitecture and Estimated Bone Strength. Kyle Kenji Stephen Nishiyama A THESIS

Transcription

1 UNIVERSITY OF CALGARY In Vivo Assessment of Bone Microarchitecture and Estimated Bone Strength by Kyle Kenji Stephen Nishiyama A THESIS SUBMITTED TO THE FACULTY OF GRADUATE STUDIES IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY BIOMEDICAL ENGINEERING GRADUATE PROGRAM CALGARY, ALBERTA OCTOBER, 2012 Kyle Kenji Stephen Nishiyama 2012

2 ABSTRACT Osteoporosis is a disease characterized by loss of bone mass and structural deterioration leading to increased risk of fracture. Currently, osteoporosis is assessed by areal bone mineral density; however, this does not provide structural information, which is a key determinant of bone strength. Recent advances allow for the assessment of bone structure in vivo using quantitative computed tomography (QCT) and high-resolution peripheral QCT (HR-pQCT). The overall objective of this thesis was to improve the assessment of bone structure and strength using three-dimensional imaging technologies. First, measurements of cortical porosity from HR-pQCT were validated against micro-ct (R 2 = 0.80) and applied to a population-based sample (N = 280, Ages: yrs.) of healthy, osteopenic, and osteoporotic, pre- and postmenopausal women. Cortical porosity was higher in postmenopausal women and those with disease. Measurements of cortical porosity were also applied to another group with high fracture incidence: children and adolescents (N = 398, Ages: 9-22 yrs.). Boys were found to have higher porosity than girls, and those at earlier pubertal stages had higher porosity than those post-pubertal. Bone quality measurements were also combined with finite element estimates of bone strength to determine if the measurements could distinguish women with fracture from fracturefree controls. High accuracy was achieved using both HR-pQCT scans of peripheral sites (83.3%) and QCT scans of the proximal femur (84.3%) when classifying the groups using support vector machines. Together, these results provide insight into the differences in bone microstructure and strength with age and disease. In addition, this work demonstrates the ability of novel 3D technologies and methods to better discriminate individuals with and without fracture. ii

3 PREFACE The following contributions were completed during the course of this thesis. This includes both works directly related to this thesis and other collaborative work. Journal Publications: Nishiyama KK, Campbell GM, Klinck RJ, Boyd SK (2010) Reproducibility of bone microarchitecture measurements in rodents by in vivo micro-computed tomography is maximized with 3D image registration. Bone. 46(1): Nishiyama KK, Macdonald HM, Buie HR, Hanley DA, Boyd SK (2010) Postmenopausal women with osteopenia have higher cortical porosity and thinner cortices at the distal radius and tibia than women with normal abmd: an in vivo HR-pQCT study. J Bone Miner Res. 25(4): Manske SL, Macdonald HM, Nishiyama KK, Boyd SK, McKay HA (2010) Clinical Tools to Evaluate Bone Strength. Clin Rev Bone Miner Metab. 8(3): Macdonald HM, Nishiyama KK, Hanley DA, Boyd SK (2011) Changes in trabecular and cortical bone microarchitecture at peripheral sites associated with 18-months of teriparatide therapy in postmenopausal women with osteoporosis. Osteoporosis Int. 22(1): Macdonald HM, Nishiyama KK, Kang J, Hanley DA, Boyd SK (2011) Age-related patterns of trabecular and cortical bone loss differ between sexes and skeletal sites: A population-based HR-pQCT study. J Bone Miner Res. 26(1): Nishiyama, KK and Boyd, SK (2011) In vivo assessment of trabecular and cortical bone microstructure. Clinical Calcium. 21(7): Nishiyama, KK, Macdonald, HM, Moore, SA, Fung, T, Boyd, SK, McKay, HA (2012) Cortical porosity is higher in boys compared with girls at the distal radius and distal tibia during pubertal growth: An HR-pQCT Study. J Bone Miner Res. 27(2): Nishiyama KK, Macdonald HM, Hanley DA, Boyd SK (Accepted) Women with previous fragility fractures can be classified based on bone microarchitecture and finite element analysis measured with HR-pQCT. Osteoporos Int. Nishiyama KK, Gilchrist S, Guy P, Cripton P, Boyd SK (Submitted) Proximal femur bone strength estimated by a computationally fast finite element analysis in a sideways fall configuration. Nishiyama KK, Ito M, Boyd SK (In Prep) Classification of women with and without hip fracture based on quantitative computed tomography and finite element analysis. iii

4 Conference Abstracts: Nishiyama, K.K., McErlain, D.D., Sandino, C., Boyd, S.K. (2012)(Poster Presentation) Effect of various boundary conditions on proximal femur finite element models. European Society of Biomechanics. Lisbon, Portugal. McErlain, D.D., Nishiyama, K.K., Sandino, C., Boyd, S.K. (2012)(Oral Presentation) Evaluation of the effect of CT image resolution on voxel based, subject specific FEA models. European Society of Biomechanics. Lisbon, Portugal. Nishiyama, K.K., Macdonald, H.M., Hanley, D.A., Boyd, S.K. (2012)(Oral Presentation) HRpQCT measured bone microarchitecture and estimated bone strength can classify postmenopausal women with and without previous forearm fractures. Canadian Orthopaedic Association Annual Meeting Ottawa, ON. Nishiyama, K.K., McErlain, D.D., Sandino, C., Boyd, S.K. (2012)(Poster Presentation) Distal radius bone strength estimated by finite element analysis based on QCT images. International Osteoporosis Foundation European Congress on Osteoporosis and Osteoarthritis. Bordeaux, France. Nishiyama, K.K., Macdonald, H.M., Hanley, D.A., Boyd, S.K. (2011)(Poster Presentation) Bone microarchitecture and finite element analysis measured with HR-pQCT can identify postmenopausal women with and without forearm fragility fractures. American Society for Bone and Mineral Research. San Diego, CA. Libanati, C., Boyd S.K., Nishiyama, K.K., Zebaze, R.M., Hanley, D.A., Zanchetta, J.R., Thomas, T., Boutroy, S., Bogado, C., Austin, M., Seeman, E. (2011)(Oral Presentation) Denosumab Decreases Cortical Porosity in Postmenopausal Women with Low Bone Mineral Density. American College of Rheumatology Annual Meeting. Chicago, IL. Zebaze, R.M., Boyd S.K., Nishiyama, K.K., Hanley, D.A., Zanchetta, J.R., Thomas, T., Boutroy, S., Bogado, C., Austin, M., Libanati, C., Seeman, E. (2011)(Oral Presentation) Denosumab Decreases Cortical Porosity in Postmenopausal Women with low bone mineral density. IOF Regionals and ANZBMS Annual Meeting and JSBMR. Gold Coast, Austrialia. Boyd S.K., Nishiyama, K.K., Zebaze, R.M., Hanley, D.A., Zanchetta, J.R., Thomas, T., Boutroy, S., Bogado, C., Austin, M., Libanati, C., Seeman, E. (2011)(Oral Presentation) Denosumab Decreases Cortical Porosity in Postmenopausal Women with Low BMD. The Endocrine Society s 93rd Annual Meeting & Expo. Boston, MA. Endocr Rev 32: OR29-4. Boyd S.K., Nishiyama, K.K., Zebaze, R.M., Hanley, D.A., Zanchetta, J.R., Thomas, T., Boutroy, S., Bogado, C., Austin, M., Libanati, C., Seeman, E. (2011)(Poster Presentation) Denosumab Decreases Cortical Porosity in Postmenopausal Women with Low BMD. 3 rd Joint Meeting of the European Calcified Tissue Society and the International Bone and Mineral Society. Athens, Greece. iv

5 Nishiyama, K.K., Macdonald, H.M., Hanley, D.A., Boyd, S.K. (2010)(Oral Presentation) Bone microarchitecture at the distal radius to predict fracture using in vivo HR-pQCT. 11th Alberta Biomedical Engineering Conference, Banff, Canada. Page 8. Nishiyama, K.K., Macdonald, H.M., Burrows, M., McKay, H.A., Boyd, S.K. (2010)(Oral Presentation) Increased Cortical Porosity during Puberty May Increase Distal Radius Fracture Risk in Boys: A HR-pQCT Study. American Society for Bone and Mineral Research. Toronto ON. J Bone Miner Res 25(S1). AbstractDetail.aspx?aid=59dc382f-5af2-4a73-ba65-63e5af5f6d9b Macdonald, H.M., Nishiyama, K.K., Kang, J., Hanley, D.A., Boyd, S.K. (2010)(Oral Presentation) Predicted Age-related Changes in Bone Microarchitecture and Strength of the Distal Radius Help to Explain Sex Differences in Forearm Fracture Risk. American Society for Bone and Mineral Research. Toronto ON. J Bone Miner Res 25(S1). AbstractDetail.aspx?aid=6ad6eb c2-809a-dc09ec02434c Nishiyama, K.K., Macdonald, H.M., Hanley, D.A., Boyd, S.K. (2010)(Poster Presentation) Trabecular bone density and microarchitecture at the distal radius are predictors of fragility fractures in postmenopausal women as measured by HR-pQCT. International Osteoporosis Foundation World Congress on Osteoporosis, Florence, Italy. Osteoporosis International. 21(S1):S223. Nishiyama, K.K., Macdonald, H.M., Hanley, D.A., Boyd, S.K. (2009)(Oral Presentation) The Influence of Cortical Porosity on Bone Strength Measured in vivo by HR-pQCT. 10th Alberta Biomedical Engineering Conference, Banff, Canada. Page 18. Nishiyama, K.K., Macdonald, H.M., Buie, H.R., Hanley, D.A., Boyd, S.K. (2009)(Poster Presentation) In vivo HR-pQCT assessment of the pattern of age-related changes in human radius cortical porosity. J Bone Miner Res 24 (S1). Accessed Sept. 29, Available at: Meeting/AbstractDetail.aspx?aid=6125ed7d-91ac- 4bfe-a4c7-646f4c2d1d4f. Macdonald, H.M., Nishiyama, K.K., Buie, H.R., Hanley, D.A., Boyd, S.K. (2009)(Poster Presentation) High parity is associated with increased cortical porosity and reduced cortical density of the distal radius as measured with high-resolution pqct. J Bone Miner Res 24 (S1). Accessed Sept. 29, Available at: 9a2bb422-a74c- 4b3e-ba36-aa22458bb59a. Macdonald, H.M., Nishiyama, K.K., Buie, H.R., Hanley, D.A., Boyd, S.K. (2009)(Poster Presentation) Parity is associated with increased cortical porosity and reduced cortical density of the distal radius as measured with high-resolution pqct. 91st Endocrine Society s Annual Meeting, Washington, DC. v

6 Nishiyama, K.K., Macdonald, H.M., Buie, H.R., Hanley, D.A., Boyd, S.K. (2009)(Poster Presentation) A population-based comparison of cortical bone thickness and porosity in preand postmenopausal women by HR-pQCT. 36th European Symposium on Calcified Tissues, Vienna, Austria. Bone 44(S2):S372-S373. Macdonald, H.M., Nishiyama, K.K., Buie, H.R., Hanley, D.A., Boyd, S.K. (2009)(Poster Presentation) 18-Months of teriparatide therapy increases both trabecular number and cortical porosity of the distal radius without reducing bone strength in postmenopausal women with osteoporosis. 36th European Symposium on Calcified Tissues, Vienna, Austria. Bone 44(S2):S428. Nishiyama, K.K., Buie, H.R., Macdonald, H.M., Hanley, D.A., Boyd, S.K. (2009)(Oral Presentation) Automated in vivo quantification of human cortical bone thickness and porosity by HR-pQCT. 32nd Conference of the Canadian Medical and Biological Engineering Society, Calgary, Alberta. Page 21. Nishiyama, K.K., Macdonald, H.M., Buie, H.R., Hanley, D.A., Boyd, S.K. (2009)(Oral Presentation) In vivo quantification of age-related changes in human cortical bone thickness and porosity by HR-pQCT. 2nd Joint Meeting of the International Bone & Mineral Society and the Australian & New Zealand Bone & Mineral Society, Sydney, Australia. Bone 44(S1):S22. Nishiyama, K.K., Macdonald, H.M., Buie, H.R., Hanley, D.A., Boyd, S.K. (2009)(Poster Presentation) Comparison of cortical bone thickness, porosity, and density in pre- and postmenopausal women measured by HR-pQCT at the distal radius and tibia. 2nd Joint Meeting of the International Bone & Mineral Society and the Australian & New Zealand Bone & Mineral Society, Sydney, Australia. Bone 44(S1):S96-S97. Nishiyama, K.K., Campbell, G.M., Klinck, R.J., Boyd, S.K. (2008)(Oral Presentation) Image registration to achieve improved reproducibility for in vivo micro-computed tomography bone quality measurements. 9th Alberta Biomedical Engineering Conference, Banff, Canada. Page 19. vi

7 ACKNOWLEDGEMENTS This work would not have been possible without the assistance of many people and organizations. First and foremost, I would like to thank my supervisor Dr. Steven Boyd whose involvement and guidance was instrumental in this work. I could not imagine a better supervisor. Also, I would like to thank my supervisory committee members Dr. David Hanley and Dr. Ross Mitchell who were always available to provide guidance and assistance. Collaborations were a key part of this work, and for this I would like to thank Dr. Masako Ito, Dr. Heather McKay, and Dr. Peter Cripton for helping me as if I was one of their own students. My unofficial supervisor Dr. Heather Macdonald spent countless hours mentoring me and reviewing my work. I am extremely grateful for her friendship and support. I am also grateful for all of the support from members of the Bone Imaging Laboratory, many of whom I consider both colleagues and friends. Specifically, Dr. Erika Kristensen for her advice and necessary laughter; Dr. Yves Pauchard for thought-provoking discussions and guidance; and Dr. Sarah Manske, Dr. Graeme Campbell, Dr. Anna-Maria Liphardt, John Schipilow, Katharina Schnakenburg, Dr. Dave McErlain, Dr. Clara Sandino, and Dr. Lauren Burt for your insight and camaraderie. I would also like to acknowledge my sources of funding: Vanier Canada Graduate Scholarships, Osteoporosis Canada and the Canadian Multicenter Osteoporosis Study, Alberta Innovates Health Solutions, Japan Society for the Promotion of Science, the University of Calgary, and Dr. Steven Boyd. Without this support none of this work would have been possible. Thank you to my parents who have always inspired and encouraged me in all of my endeavours. Everything I have done is a result of their guidance and support. Also, thank you to my brothers for all of the healthy competition over the years. vii

8 The Human Performance Lab provided an excellent environment both learning and building friendships. Thanks to all my friends, especially Erik Groves, Brandon Hisey, Stephen Andrews, and Andrew Betik for the great memories. Finally, thanks to Meaghan Nolan, who has encouraged and supported me throughout my degree. She has had an amazing impact on my life and work, and has always been there for me. viii

9 TABLE OF CONTENTS ABSTRACT... ii PREFACE... iii ACKNOWLEDGEMENTS... vii TABLE OF CONTENTS... ix LIST OF TABLES... xii LIST OF FIGURES... xiv LIST OF ABBREVIATIONS... xviii CHAPTER ONE: INTRODUCTION Motivation Objectives and Hypotheses... 4 Objective One... 4 Objective Two Outline of Thesis... 6 CHAPTER TWO: BACKGROUND AND LITERATURE REVIEW Bone Cortical Bone Structure Trabecular Bone Structure Bone Cells Bone Growth and Acquisition Bone Modeling, Remodeling, and Adaptation Imbalances Resulting in a High Risk of Fracture Medical Imaging Techniques Dual-energy X-ray Absorptiometry Computed Tomography High-Resolution Peripheral Quantitative Computed Tomography Artefacts and Limitations of CT Quantitative Image Evaluation Image Segmentation Image Calibration HR-pQCT Morphometric Indices CT Measurements and Fracture Risk Bone Mechanics Basic Mechanics Bone Mechanical Properties Site-Specific Bone Mechanical Testing Finite Element Analysis Image-Based Finite Element Analysis Bone Strength Estimates and Fracture Machine Learning Classification Association Models Classification Models Machine Learning Methods Knowledge Gaps and Summary ix

10 CHAPTER THREE: POSTMENOPAUSAL WOMEN WITH OSTEOPENIA HAVE HIGHER CORTICAL POROSITY AND THINNER CORTICES AT THE DISTAL RADIUS THAN WOMEN WITH NORMAL ABMD: AN IN VIVO HR-PQCT STUDY Introduction Materials and Methods In Vitro Validation Specimens HR-pQCT and µct Scanning Image Registration and Processing Cortical Bone Measurements In Vivo Assessment Subjects In Vivo HR-pQCT Scanning Statistical Analysis Results In Vitro Validation In Vivo Assessment Discussion Validation of HR-pQCT for Cortical Porosity and Thickness Comparison of Pre- and Postmenopausal, Normal, Osteopenic, and Osteoporotic Women CHAPTER FOUR: CORTICAL POROSITY IS HIGHER IN BOYS COMPARED WITH GIRLS AT THE DISTAL RADIUS AND DISTAL TIBIA DURING PUBERTAL GROWTH: AN HR-PQCT STUDY Introduction Materials and Methods Participants HR-pQCT Scan Acquisition Image Processing and Measurements Finite Element Analysis Statistical Analysis Results Participant Descriptives Comparison Between Sexes Within Maturity Categories Comparison Across Maturity Categories Within Sex Discussion CHAPTER FIVE: WOMEN WITH PREVIOUS FRAGILITY FRACTURES CAN BE CLASSIFIED BASED ON BONE MICROARCHITECTURE AND FINITE ELEMENT ANALYSIS MEASURED WITH HR-PQCT Introduction Materials and Methods Participants HR-pQCT Scan Acquisition and Measurements Finite Element Analysis x

11 5.2.4 Statistical Analysis Results Participant Characteristics and Bone Outcome Variables SVM Classification Discussion CHAPTER SIX: PROXIMAL FEMUR BONE STRENGTH ESTIMATED BY A COMPUTATIONALLY FAST FINITE ELEMENT ANALYSIS IN A SIDEWAYS FALL CONFIGURATION Introduction Materials and Methods Specimens CT Scan Acquisition Mechanical Testing Image Processing Finite Element Analysis Statistical Analysis Results Discussion CHAPTER SEVEN: CLASSIFICATION OF WOMEN WITH AND WITHOUT HIP FRACTURE BASED ON QUANTITATIVE COMPUTED TOMOGRAPHY AND FINITE ELEMENT ANALYSIS Introduction Materials and Methods Participants CT Scan Acquisition Image Processing Finite Element Analysis Statistical Analysis Results Discussion CHAPTER EIGHT: DISCUSSION, FUTURE WORK, AND CONCLUSIONS Discussion What Does Cortical Porosity Tell Us? Can We Predict Who Will Fracture? Future Work Conclusions xi

12 LIST OF TABLES Table 2.1: Comparison of in vivo HR-pQCT and QCT scanning Table 2.2: Morphological indices used to quantify bone structure with HR-pQCT Table 2.3: Various density (ρ) to modulus (E) relationships reported in previous studies. R 2 is the determination coefficient. A more comprehensive list can be can be found in (Helgason et al., 2008) Table 3.1: Regression analysis results between µct and HR-pQCT measures, all values p< Table 3.2: Descriptive characteristics of the subjects by group (mean ± SD) Table 3.3: Outcome mean ± SD and percent differences between groups. Pre- refers to premenopausal and Post- refers to postmenopausal. a-c. Significant difference between groups: a. P<0.001, b. P<0.01, c. P<0.05, NS, Not significant with Bonferroni correction Table 4.1: Participant descriptives by puberty group for girls and boys [Mean (SD)]. Note: All p values after Bonferroni correction. a. p<0.001, b. p<0.01, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex Table 4.2: Bone microstructure at the distal radius across puberty groups for girls and boys [Mean (SD)]. All p-values after Bonferroni correction. a. p<0.001, b. p<0.01, c. p<0.05, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex Table 4.3: Bone microstructure at the distal tibia across puberty groups for girls and boys [Mean (SD)]. All p-values after Bonferroni correction. a. p<0.001, b. p<0.01, c. p<0.05, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex Table 4.4: Finite element estimated bone strength parameters at the distal radius and distal tibia across puberty groups for girls and boys [Mean (SD)]. All p-values after Bonferroni correction. a. p<0.001, b. p<0.01, c. p<0.05, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex Table 5.1: Descriptive characteristics of the postmenopausal women with forearm fractures and fractures at all sites, and their corresponding age-matched controls Table 5.2: Bone microarchitecture and finite element parameters [mean (SD)] for the fracture and control groups at the distal radius xii

13 Table 5.3: Bone microarchitecture and finite element parameters [mean (SD)] for the fracture and control groups at the distal tibia Table 5.4: DXA parameters [mean (SD)] for the fracture and control groups Table 5.5: SVM classification results for the forearm fractures and for fractures at all sites. Five different models were used for each fracture group. The first was HR-pQCT, FE, and all DXA parameters, the second was HR-pQCT and FE parameters, the third was FE parameters alone, the fourth was all DXA parameters including simulated ultra-distal radius (UDR), and the last was lumbar spine (LS) and femoral neck (FN) DXA only. Accuracy, sensitivity, specificity, and the receiver operating characteristic area under the curve (ROC-AUC) are shown for each model Table 6.1: Age, sex, weight, height and total femoral neck (FN) abmd measured by DXA for the cadaver proximal femur specimens Table 7.1: Descriptive characteristics as Mean ± SD for all fractures versus controls then broken down into trochanteric fractures versus controls, and neck fractures versus controls. P-values represent differences between the fracture groups and their corresponding controls Table 7.2: FE estimates of stiffness, failure load and QCT-vBMD. Results are grouped by all fractures versus controls then broken down by trochanteric fractures versus controls, and neck fractures versus controls. a. p<0.001, b. p<0.01, c. p<0.05 between the fracture groups and their corresponding controls after Bonferroni adjustment Table 7.3: Accuracy, sensitivity, specificity, and receiver operating characteristic areas under the curve (ROC-AUC) for the SVM models. The first are based on both types of fractures pooled together, the second on trochanteric fractures, and the third only on femoral neck fractures xiii

14 LIST OF FIGURES Figure 2.1: Schematic diagram of cortical and trabecular bone showing hierarchical levels of organization. Adapted from (Weatherholt et al., 2011) with permission Figure 2.2: Bone cells and their relationships. The long processes in the osteocytes sense a microcrack and signal osteoclasts to resorb the bone matrix. Osteoblasts follow behind deposit new bone. Reproduced with permission from (Seeman and Delmas, 2006). Copyright Massachusetts Medical Society Figure 2.3: Schematic of endochondral ossification. Chondrocytes form a cartilage model and a primary ossification center is formed. Next, blood vessels penetrate the cavity and secondary ossification centers are created. From (Wojcicka et al., 2012) with permission Figure 2.4: Peak bone mass accumulation for boys (solid line) and girls (dashed line). The peak accumulation occurs later for boys compared with girls and boys have a greater maximum accumulation during growth. Adapted from (Bailey et al., 1999) with permission Figure 2.5: The mechanostat theory demonstrating how bone adapts to its mechanical environment. During times of low strain there is a reduction of bone mass through remodelling and during times of high strains there is increased bone formation. Adapted from (Frost, 1997) with permission Figure 2.6: Example proximal femur DXA scan. Lines show the boundaries used to measure abmd Figure 2.7: CT backprojection reconstruction. The left panel shows the object to be imaged and a representation of the attenuation values measured by the detector in the line below. The attenuation values are smeared back across the image for one (center) and two (right) projections. As more projections are included, the image (a circle) becomes clearer Figure 2.8: Sample axial (left) and coronal (center) image slices from a hip QCT scan. A three-dimensional reconstructed image from a QCT scan is shown on the far right Figure 2.9: Scanning with HR-pQCT (XtremeCT, Scanco Medical, Brüttisellen, Switzerland) at the distal tibia Figure 2.10: HR-pQCT images of the distal radius (left) and distal tibia (right). Images slices are acquired at an 82 μm isotropic resolution and can be used to form a threedimensional image Figure 2.11: HR-pQCT distal radius image with the cortical region segmented and shown in transparent gray. The cortical pores are highlighted in blue xiv

15 Figure 2.12: CT image artefacts. The left image depicts partial volume effects when trying to represent the red circle. The image pixels that overlap the border of the object are neither dark gray nor white but a proportion of both. The right image shows a HR-pQCT scan with motion artefacts seen distinctly as streaks and breaks in the cortex Figure 2.13: Result of an automatic segmentation by the dual-threshold method of the cortical and trabecular regions in a distal radius HR-pQCT image. The cortex is shown in blue and the segmented trabecular region is shown in gray Figure 2.14: Sample hip CT scan with calibration phantom underneath the subject (left). Enlarged portion shows the phantom used to calibrate the density measurements. On the right is an example plot to determine the linearity of the scan as well as the values to calibrate the image Figure 2.15: Load-deformation plot (left) and stress-strain plot (right) depicting the various properties that can be used to characterize a material Figure 2.16: Example load-deformation plots for different diseases showing varying stiffness as well as proportions of elastic and plastic regions. Adapted with permission from (Sato et al., 1999). Copyright American Chemical Society Figure 2.17: Falling (left) and standing (right) loading configurations for the mechanical testing of the proximal femur. Force arrows are indicated in red Figure 2.18: An HR-pQCT tibia scan with 1% uniaxial compression applied (far left). Other images showing various views of the FE model coloured by Von Mises stress (MPa) Figure 2.19: Example of a simple two-dimensional support vector machine classification. The dashed line shows the optimal hyperplane margin to separate the two classes (squares and circles) and the support vectors are shown in red. Adapted from (Cortes & Vapnik, 1995) with permission Figure 3.1: Scout views representing the reference line (solid line) and region scanned (dotted lines) at the distal radius (A.) and distal tibia (B.) Figure 3.2: Comparison of cortical bone measurements from HR-pQCT with the gold standard µct. Regression analyses are shown on the left and Bland-Altman plots on the right for cortical porosity (Ct.Po), cortical thickness (Ct.Th), number of pores, mean volume of pores, and cortical density. Solid lines indicate 95% confidence intervals Figure 3.3: Three-dimensional images of the same region scanned with µct (A.) and HRpQCT (B.) visually showing the resolution differences. The voxels sizes are 19µm and 82µm, respectively Figure 3.4: Single slice and three-dimensional representative images of the original grayscale image (left), the segmented images from the auto-segmentation method (center) and threshold-filter method (right). The regions designated as trabecular bone are shown in gray, cortical bone is shown in blue, and cortical pores are shown in yellow xv

16 Figure 4.1: Numbers of participants at baseline recruited from each cohort and number of follow-up scans analyzed for boys and girls at the distal radius and distal tibia Figure 4.2: Scout view images of the distal radius (A) and distal tibia (B) illustrating the 7% and 8% measurement sites, respectively Figure 4.3: Plots of cortical density (Ct.BMD), cortical porosity (Ct.Po), cortical area (Ct.Ar), and failure load at the distal radius for girls and boys by puberty group. Error bars represent SE. a. p<0.001, b. p<0.01, c. p<0.05: significant difference between girls and boys within the same puberty group. d. p<0.001 e. p<0.01: significant difference between puberty group and the PRE group within sex. All p-values are after Bonferroni correction Figure 4.4: Plots of cortical density (Ct.BMD) and cortical porosity (Ct.Po) at the distal tibia for girls and boys by puberty group. Error bars represent SE. a. p<0.001, b. p<0.01: significant difference between girls and boys within the same puberty group. d. p<0.001 e. p<0.01: significant difference between puberty group and the PRE group within sex. All p-values are after Bonferroni correction Figure 4.5 A schematic representation of differences in total bone size and cortical bone density for girls (G) and boys (B) across puberty (assessed using the method of Tanner (T)). For our purposes we defined Tanner stage 1 as pre-puberty (PRE), Tanner stages 2 and 3 as early puberty (EARLY), Tanner stage 4 as peri-puberty (PERI) and Tanner stage 5 as post-puberty (POST). Significant differences between girls and boys are shown for finite element estimated failure load (where boys values exceed girls after early puberty) and cortical porosity (Ct.Po) where boys values exceed girls after pre puberty. (Diagram not exact scale) Figure 5.1: Representative distal radius scans of a participant with a low trauma forearm fracture (left) and an age-matched non-fracture control (right) Figure 6.1: Photograph of the testing apparatus and specimen positioned in the mechanical testing device (top) and corresponding schematic of the finite element model boundary conditions where a 1 mm displacement is applied to the PMMA cap (represented in white) on the greater trochanter (bottom) Figure 6.2: Comparison of experimental versus the finite element estimates of stiffness (A) and failure load (B). Bland-Altman plots depict the mean of the experimental and FE estimated stiffness (C) versus the difference between the two values and the corresponding plot for failure load (D). The horizontal lines indicate the mean values and 95% confidence intervals Figure 6.3: The overall FE estimated structural stiffness in the loading direction with the femoral neck internally rotated -30, -15, 0, 15, 30, and 45. After a Bonferroni adjustment, significantly (p<0.05) higher stiffness values are indicated: a) compared with 45, b) compared with 30, c) compared with 15, and d) compared with -30. Error bars represent standard error xvi

17 Figure 7.1: Schematic of the angles of rotation for the femoral neck used for estimating the bone strength. A left femur is shown with the femoral neck internal rotation angle ranging from -30 to 45 at 15 intervals Figure 8.1: Cortical porosity in women (left) and men (right) at the distal radius (top row) and distal tibia (bottom row). Porosity (%) is plotted against age in years. Adapted from (Macdonald et al., 2011) with permission xvii

18 LIST OF ABBREVIATIONS Abbreviation DXA QCT Micro-CT HR-pQCT MR abmd vbmd Tb.BMD Ct.BMD BV/TV Tb.N Tb.Th Tb.Sp Ct.Th Ct.Po Ct.Ar Tb.Ar Tt.Ar HA ROI OR TN TP FP FN ROC AUC SVM Definition Dual-energy X-ray Absorptiometry Quantitative Computed Tomography Micro Computed Tomography High-Resolution Peripheral Quantitative Computed Tomography Magnetic Resonance Areal Bone Mineral Density Volumetric Bone Mineral Density Trabecular Bone Mineral Density Cortical Bone Mineral Density Bone Volume Fraction (Bone Volume / Total Volume) Trabecular Number Trabecular Thickness Trabecular Separation Cortical Thickness Cortical Porosity Cortical Cross-sectional Area Trabecular Cross-sectional Area Total Cross-sectional Area Hydroxyapatite Region of Interest Odds Ratio True Negative True Positive False Positive False Negative Receiver Operating Characteristic Area Under the Curve Support Vector Machine xviii

19 CHAPTER ONE Introduction 1

20 CHAPTER ONE: INTRODUCTION 1.1 Motivation Osteoporosis is a serious degenerative disease characterized by both the loss of bone mass and the deterioration of bone architecture (NIH Consensus Development Panel on Osteoporosis Prevention, 2001). At least 1 in 3 women and 1 in 5 men will suffer an osteoporotic fracture within their lifetime (Cummings et al., 1989). Osteoporotic fractures are often debilitating, significantly reducing an individual s quality of life, and their treatment poses a significant burden on the health care system, costing approximately $1.9 billion per year in Canada alone (Wiktorowicz et al., 2001). Current clinical diagnosis of osteoporosis relies on measurements of areal bone mineral density (abmd) by dual-energy x-ray absorptiometry (DXA). However, due to DXA s twodimensional nature, these measurements are affected by the size and position of the subject (Bolotin, 2007). Epidemiological data indicate that over 80% of fractures occur in women who would not be classified as osteoporotic according to current abmd criteria (Stone et al., 2003), highlighting the limitations of this approach and the need for better assessment methods. One of the reasons DXA-measured abmd is such a poor predictor of fracture risk is that it fails to provide any structural information about bone. Bone structure is fundamental to bone strength, and it is hypothesized that osteoporosis-related changes in bone architecture lead to decreased bone strength and increased fracture risk. Thus, in vivo structural information is critical for the assessment of disease progression, treatment effects, and risk of fracture (Müller and Rüegsegger, 1995). 2

21 Recent advances in medical imaging techniques allow bone microstructure to be assessed in vivo non-invasively. At various skeletal sites within the body, three-dimensional macro- and microstructure of both cortical and trabecular regions can be imaged. The primary tools for assessing volumetric density and microstructure are quantitative computed tomography (QCT) and more recently, high-resolution peripheral quantitative computed tomography (HR-pQCT). Using the three-dimensional images produced by HR-pQCT and QCT, finite element (FE) analysis can be used to generate patient-specific estimates of bone strength. FE models using HR-pQCT images have been shown to better predict fractures than HR-pQCT-measured BMD or microstructure alone (Boutroy et al., 2008), and FE models using QCT images have been shown to better predict femoral strength than QCT-based vbmd and DXA-based abmd (Cody et al., 1999; Crawford et al., 2003). 3

22 1.2 Objectives and Hypotheses The overall goal of this thesis is to use novel imaging techniques and analysis methods to improve the assessment of bone quality and fracture risk. This goal involves two main objectives, the first focused on using HR-pQCT and the second using QCT. Objective One Cortical bone is a major contributor to whole bone strength at both the distal radius and distal tibia (MacNeil and Boyd, 2007b; Spadaro et al., 1994); therefore, determining the bone quality and strength of the cortex is important for the prediction of fracture risk. Cortical porosity, which describes the canals that exist in bone s outer shell, provides space for vessels and nerves. Cortical porosity has been shown to increase with age and disease (Bousson et al., 2001; Stein et al., 1999a), and to be a predictor of bone strength (McCalden et al., 1993; Wachter et al., 2001a; Yeni et al., 1997). Attempts have been made to use DXA to predict cortical bone strength; however, correlations have been weak likely because the detailed architecture of the cortex could not be resolved (Snyder and Schneider, 1991). Until now, in vivo quantification of cortical porosity has been limited by the high-resolution required to resolve the pores and difficulties involved in differentiating the cortical and trabecular regions. Increased cortical porosity is associated with high rates of bone turnover that can occur during menopause and disease (Brockstedt et al., 1993). It is estimated that age-related increases in porosity account for 76% of the loss in bone strength in vitro at the proximal femur (McCalden et al., 1993). Growth and puberty is another time when it has been hypothesized that high cortical porosity exists due to the high turnover and rapid growth (Parfitt, 1994b). 4

23 The first objective of this thesis is to develop validated methods to measure cortical porosity in vivo and determine age-, sex-, and disease-related differences in adult and adolescent populations. Specific Hypotheses: (1.1) Cortical porosity is higher in postmenopausal women compared with premenopausal women and is higher in those with osteopenia and osteoporosis compared with those who have normal abmd. (1.2) Cortical porosity is higher in boys compared with girls during puberty and is highest in pubertal boys compared with prepubertal and postpubertal boys. (1.3) HR-pQCT measurements of bone quality, including cortical porosity, and estimated bone strength will be able to accurately distinguish between women with and without low-trauma fractures. Objective Two Clinical QCT is capable of providing structural measurements and volumetric density measurements at the proximal femur and lumbar spine, the most common fragility fracture sites within the body (Cummings et al., 1989). When these images are combined with finite element (FE) analysis it has been shown to be a promising technique to estimate bone strength at the proximal femur (Cody et al., 1999; Dragomir-Daescu et al., 2011; Keyak et al., 2001; Koivumäki et al., 2012). However, FE analysis is limited by the large manual component required to prepare the model as well as the high computational cost and time involved in solving the FE models 5

24 (Keyak et al., 2001; Koivumäki et al., 2012). In addition, since it is unknown how a subject will fall, it is necessary to incorporate various loading configurations when predicting bone strength. The second objective of this thesis is to develop a validated tool to provide a noninvasive estimate of femoral bone strength and apply the tool to a clinical dataset of women who have suffered a fracture and fracture-free, age-matched controls. Specific Hypotheses: (2.1) Estimates of bone failure load and stiffness from FE analysis will have a high agreement with estimates from mechanical testing of cadaver femurs. (2.2) Since it is not possible to predict how a specific subject will fall, estimating bone strength in a test space of loading conditions that encompasses various possible fall forces will allow for an accurate classification of women with and without fracture. 1.3 Outline of Thesis Chapter Two summarizes background information and current literature related to this thesis. This begins with a review of bone biology, followed by medical imaging techniques, mechanics, and data classification methods. Chapter Three includes a validation of cortical porosity measurements by HR-pQCT and compares cortical porosity between pre- and postmenopausal women as well as those with osteopenia and osteoporosis. This chapter is based on a manuscript published in the Journal of Bone and Mineral Research. 6

25 Chapter Four extends the methods developed in Chapter Three and applies them to children and adolescents during growth and bone acquisition. This chapter is also based on a published manuscript in the Journal of Bone and Mineral Research. Chapter Five analyzes bone microarchitecture and estimated bone strength in women who have suffered a low-trauma fracture as well as in age-matched, fracture-free women in order to determine if support vector machines can distinguish between the two groups. This work has been accepted as an original manuscript for publication in Osteoporosis International. Chapter Six is based on a technical study to validate a method for finite element analysis estimates of bone strength from QCT images. It also includes a sensitivity analysis of various loading configurations. This work has been submitted as an original manuscript. Chapter Seven applies the methods developed in Chapter Six to a cohort of women who have suffered femoral neck fractures, trochanteric fractures, and age-matched controls. The estimates of bone strength are used to distinguish the women with fracture from the control women. This work is in preparation for submission as a manuscript. Finally, Chapter Eight includes an integrated discussion of the studies presented to improve the in vivo assessment of bone quality and fracture risk, future proposed work, and final conclusions. 7

26 CHAPTER TWO Background and Literature Review 8

27 CHAPTER TWO: BACKGROUND AND LITERATURE REVIEW 2.1 Bone Bone is a complex living tissue that serves mechanical support and protective functions as well as provides metabolic functions within the body. Bone is composed of both organic and inorganic materials. The organic portion is mainly composed of collagen (Type 1), which forms an extracellular matrix where the inorganic portion can be deposited to strengthen the structure. These inorganic deposits are mainly composed of hydroxyapatite. At the macroscopic level, bone has two compartments: cortical, or compact bone, and trabecular, or cancellous bone (Cowin, 2001) Cortical Bone Structure Cortical bone is densely packed bone and is mainly present in the shafts of long bones such as the femur and tibia. This compartment makes up approximately 80% of the bone mass in the whole skeleton, yet it accounts for a much smaller surface area than trabecular bone (Bronner and Worrell, 1999). The cortex is comprised of three hierarchical levels of structural organization (Figure 2.1). The first level has an approximate size of 100 to 300 μm and is composed mainly of primary and secondary osteons and interstitial bone (Martin and Burr, 1989). Osteons are cylindrical structures surrounding vascular space and are oriented with the long axis of the bone, which together are known as a Haversian system. Interstitial bone fills the spaces between these Haversian systems. At the second level there are lamellae, lacunae, and cannaliculi with an approximate size of 2 to 20 μm. Lamellae make up the concentric circles observed in osteonal bone. The lacunae and cannaliculi are the canals within the matrix that 9

28 contain bone cells. Finally, the third level (0.06 to 0.6 μm) contains the collagen fibres with inorganic mineral deposits. This level can be broken into two types: woven and lamellar bone. Woven bone is built initially with a random distribution while lamellar bone follows in adjacent layers. Figure 2.1: Schematic diagram of cortical and trabecular bone showing hierarchical levels of organization. Adapted from (Weatherholt et al., 2011) with permission. 10

29 2.1.2 Trabecular Bone Structure In general, the basic structure of trabecular bone is similar to cortical bone; however, on the macroscopic level it is less densely packed and consists of interconnecting plates and rods (Martin & Burr, 1989). Trabecular bone is mainly found in the ends of long bones as well as in the vertebrae. Similarly to cortical bone, trabecular bone can be described as having hierarchical levels. The first level of trabecular bone is the individual trabeculae that are on the order of 75 to 200 μm. At this level, trabecular bone is much more porous than cortical bone and the spaces are filled with bone marrow (Figure 2.1). Unlike cortical bone there are no central canals containing vessels in the trabecular bone. At the next hierarchical level, lamellae are arranged longitudinally with the trabecular unlike the concentric arrangement in cortical bone (Kragstrup et al., 1983). At the smallest level, the collagen and mineral composites are the same as in cortical bone Bone Cells Within both trabecular and cortical bone there are three main types of bone cells involved in regulating growth and maintaining the structure: osteoclasts, osteoblasts, and osteocytes (Figure 2.2). Osteoclasts are derived from hematopoietic stem cells and are mainly responsible for the resorption of bone. Once osteoclasts are differentiated they are capable of secreting enzymes that dissolve the bone matrix. Osteoclastic activity is regulated by RANK ligand, which is secreted by the osteoblasts. Osteoblasts are derived from mesenchymal stem cells and are responsible for building the organic bone matrix and regulating mineral deposition on the matrix (Rosen, 2009). Some osteoblasts undergo further differentiation into osteocytes, which are incorporated into the bone 11

30 matrix. Their long, slender processes form a network that sense strains and signal surface cells to engage in formation or resorption (Bonewald, 2007). Figure 2.2: Bone cells and their relationships. The long processes in the osteocytes sense a microcrack and signal osteoclasts to resorb the bone matrix. Osteoblasts follow behind deposit new bone. Reproduced with permission from (Seeman and Delmas, 2006). Copyright Massachusetts Medical Society Bone Growth and Acquisition During development, there are two main methods by which bones are formed: intramembranous ossification and endochondral ossification. The majority of cortical bone is formed by intramembranous ossification including flat bones and bones in the skull. This growth is initiated by condensation of mesenchymal tissues that differentiate into osteoblasts, eventually forming an osteoid. From the osteoid, spicules form to make trabecular bone. Cortical bone is formed when the spaces between the trabeculae are filled by osteons or Haversian systems (Cowin, 2001). 12

31 Endochondral ossification forms the majority of long bones as well as trabecular bone as shown in Figure 2.3. This method of growth is initiated by chondrocytes proliferating and depositing a matrix. Cartilage cells in this matrix calcify to form a primary ossification center. At these centers, osteoblasts secrete osteoid and trabecular bone is formed. Blood vessels penetrate the remaining cartilage and secondary ossification centers are formed. Eventually, in these secondary centers, the cartilage is reduced to thin growth plates where longitudinal bone growth occurs. Finally, at maturity, these growth plates fuse and the remainder of the cartilage is mineralized (Mackie et al., 2008). Figure 2.3: Schematic of endochondral ossification. Chondrocytes form a cartilage model and a primary ossification center is formed. Next, blood vessels penetrate the cavity and secondary ossification centers are created. From (Wojcicka et al., 2012) with permission. Through these growth mechanisms, the majority of bone mass is accumulated during childhood and adolescence (Rizzoli et al., 2001). The peak of this bone mass accumulation 13

32 Total body bone gain g/yr coincides with puberty rather than chronological age particularly when comparing girls and boys (Bailey et al., 1999). Figure 2.4 depicts boys peak bone mass accumulation around 14 years of age while girls peak occurs earlier between 12 and 13 years of age (Bailey et al., 1999). Interestingly, the time of peak bone mass accumulation also corresponds with the peak incidence of fractures (Cooper et al., 2004b; Landin, 1983). It has also been suggested that a decreased accumulation of bone mass during growth may lead to an increased risk of osteoporosis in the future (Ferrari et al., 2006; Rizzoli and Bonjour, 1999) Age in Years Figure 2.4: Peak bone mass accumulation for boys (solid line) and girls (dashed line). The peak accumulation occurs later for boys compared with girls and boys have a greater maximum accumulation during growth. Adapted from (Bailey et al., 1999) with permission. 14

33 2.1.5 Bone Modeling, Remodeling, and Adaptation Closely linked with the processes of bone growth is bone modeling, which controls the outer shape and size of the bones as well as the shape and size of the inner cavities. In this process, bone formation by the osteoblasts occurs faster than resorption by the osteoclasts over the bone surfaces. The majority of this modeling occurs during growth in the form of periosteal apposition in conjunction with endosteal resorption, resulting in bones with larger cross sectional areas and only minimal changes in cortical thickness (Duan et al., 2003; Kontulainen et al., 2006). After maturity, some periosteal bone modelling continues throughout life, particularly in men, which is observed as an increase in bone diamater with age (Seeman, 2003). In contrast, bone remodeling is a continuous process that occurs throughout life on the trabecular, endocortical, and intracortical surfaces. It is composed of both resportion and formation phases that maintain an approximately constant bone mass. By this process bone is able to adapt and repair microdamage (Parfitt, 2002). The mechanostat theory provides an explanation of how mechanical loading can cause bone modeling and remodelling as shown in Figure 2.5 (Frost, 1987). In a state of underload the bone mass will be reduced through increased resorption. In a state of overload, the bone mass will increase through increased formation. 15

34 Figure 2.5: The mechanostat theory demonstrating how bone adapts to its mechanical environment. During times of low strain there is a reduction of bone mass through remodelling and during times of high strains there is increased bone formation. Adapted from (Frost, 1997) with permission Imbalances Resulting in a High Risk of Fracture When the above-mentioned processes are modified due to accelerated growth or biological changes, imbalances in bone remodeling or bone deficiencies can cause decreased bone strength and increased risk of fracture (Ammann and Rizzoli, 2003; Wang et al., 2010). In a disease such as osteoporosis, there is a loss of bone mass and a deterioration of the bone microarchitecture. This is caused by an imbalance between bone resorption by osteoclastic activity and bone formation by osteoblasts (NIH Consensus Development Panel on Osteoporosis 16

35 Prevention, 2001). On the microstructural level, this loss of bone may include thinning and a decrease in the number of trabeculae as well as increased cortical porosity and decreased cortical thickness. While osteoporosis and increased fracture risk can affect many populations including premenopausal women (Gourlay and Brown, 2004), and men (Bilezikian, 1999), the most common observance is in postmenopausal women (Melton et al., 2009). During this time there is a decrease in the amount of estrogen in the body which causes decreased bone turnover and increased bone loss (Järvinen et al., 2003). Fractures are also a common occurrence in children and account for 10-25% of all paediatric trauma (Ferrari et al., 2006). The peak incidence of fractures occurs during the pubertal growth spurt, around the age of 14 years in boys and 11 years in girls (Cooper et al., 2004a) and approximately 42-51% of boys and 27-40% of girls will suffer a fracture during growth (Jones et al., 2002b; Landin, 1983). In the past, when boys had much higher rates of participation in recreational activities than girls, there was a higher incidence of fractures in boys (Kramhøft and Bødtker, 1988). More recently, as activity levels between boys and girls have become more similar, this difference in fracture rates has become smaller in magnitude (Cooper et al., 2004a). However, recreation or high-risk activities only partially explain fracture rates (Bailey et al., 1989; Parfitt, 1994b). It has been suggested that this high incidence of fracture could be caused by a deficiency in bone, particularly cortical bone (Wang et al., 2010), during rapid periods of growth. The importance of investigating this time of high fracture incidence is two-fold since it has also been suggested that childhood fractures may be associated with fractures or osteoporosis later in life (Ferrari et al., 2006). 17

36 2.2 Medical Imaging Techniques Medical imaging methods can provide important information regarding fracture risk. Bone strength is a key contributor to fracture risk; however, directly measuring bone strength is not possible. Bone mineral density is the most commonly used surrogate for strength, but does not take into account the structure of the bone. There is now an opportunity to use threedimensional medical imaging technologies to improve the assessment of bone strength and fracture risk. The following sections will provide an overview of medical imaging techniques that can be used to assess bone quality and strength to better understand bone biology and fracture risk described in the previous sections Dual-energy X-ray Absorptiometry The most commonly used tool for the diagnosis of osteoporosis is dual-energy x-ray absorptiometry (DXA), which allows clinicians to obtain a measure of areal bone mineral density (abmd; g/cm 2 ). DXA uses x-rays at two different energies to distinguish bone and soft tissue. It can be used in vivo at all sites within the body. DXA-measured abmd is the clinical standard as a surrogate for bone strength, and a decrease in abmd is associated with a decrease in bone strength (Bouxsein et al., 1999). Bouxsein and colleagues also indicated that abmd predicts about 66-74% of the variance in bone strength (Bouxsein et al., 1999). 18

37 Figure 2.6: Example proximal femur DXA scan. Lines show the boundaries used to measure abmd. The major limitation with DXA and abmd is that it is a two-dimensional areal measure rather than a volumetric measure of bone density (Figure 2.6). Because of this, abmd measures are affected by both the size and orientation of the bone. This limitation may explain why more than 80% of fractures occur in individuals without osteoporosis as defined by abmd alone (Stone et al., 2003). Another limitation of DXA is its inability to distinguish between cortical and trabecular bone, or reveal microarchitecture. Since bone strength is dependent on both bone quantity and architecture there is a need for other three-dimensional tools that can provide measurements of volumetric BMD and architectural characteristics. 19

38 2.2.2 Computed Tomography Computed tomography (CT) is non-destructive tool that is capable of measuring volumetric bone mineral density (vbmd; g/cm 3 ) as well as geometry and macroarchitecture in vivo at all sites within the body. It is an x-ray-based modality that obtains projections around the subject being scanned to produce a three-dimensional image (Dowsett et al., 2006). In most current in vivo CT scanners, the source and detector move around the subject and the emitted x- rays are collected from various views. As the x-rays pass through the subject, the photons are attenuated to different extents. This attenuation depends on the linear attenuation coefficient of the material through which the x-rays are passing. The projections are then reconstructed into two-dimensional slice images where the attenuation is represented as CT number in Hounsfield units (Dowsett et al., 2006). Briefly, image reconstruction of CT projections is most commonly performed using filtered backprojection. This method first filters the projections then smears the projections across the path of the x-ray as shown in Figure 2.7. When multiple projections are included, the high and low attenuation locations become apparent in the image (Dowsett et al., 2006). While not reviewed in this thesis it is important to note that there are other image reconstruction methods less commonly used in medical imaging including the Fourier transform and algebraic reconstruction (Pan et al., 2009). 20

39 Figure 2.7: CT backprojection reconstruction. The left panel shows the object to be imaged and a representation of the attenuation values measured by the detector in the line below. The attenuation values are smeared back across the image for one (center) and two (right) projections. As more projections are included, the image (a circle) becomes clearer. Current CT technology allows for scanning clinically relevant sites including the proximal femur (Figure 2.8) and lumbar spine. The in-plane resolution of these scans is typically around μm and have a slice thickness between 0.5 and 3.0 mm (Rosen, 2009). When a calibration phantom is included in the scans, computed tomography can provide quantitative measures of bone mineral density and is appropriately termed quantitative CT (QCT). This calibration process will be further described in section

40 Figure 2.8: Sample axial (left) and coronal (center) image slices from a hip QCT scan. A threedimensional reconstructed image from a QCT scan is shown on the far right High-Resolution Peripheral Quantitative Computed Tomography Recently introduced commercially in 2005, high-resolution peripheral quantitative computed tomography (HR-pQCT) has emerged as an in vivo tool to image and analyze the three-dimensional microarchitecture of bone (Boutroy et al., 2005; Laib et al., 1998). In the past, examining these material and microarchitecture properties required invasive and destructive techniques such as biopsies or cadaver studies. HR-pQCT works on the same principles as QCT; however, is limited to peripheral sites such as the radius and tibia (Figure 2.9) as this smaller field of view is the trade-off for the high-resolution. 22

41 Figure 2.9: Scanning with HR-pQCT (XtremeCT, Scanco Medical, Brüttisellen, Switzerland) at the distal tibia. Figure 2.10: HR-pQCT images of the distal radius (left) and distal tibia (right). Images slices are acquired at an 82 μm isotropic resolution and can be used to form a three-dimensional image. HR-pQCT scans allow for very high-resolution images with a nominal isotropic resolution of 82 µm (Figure 2.10) (Boutroy et al., 2005; MacNeil & Boyd, 2007b). Recent work has shown good associations between HR-pQCT measurements of distal sites with QCT 23

42 assessment of central sites (Liu et al., 2010b). A brief comparison between the two modalities is shown in Table 2.1. Table 2.1: Comparison of in vivo HR-pQCT and QCT scanning. HR-pQCT QCT Voxel size 82 μm isotropic μm in-plane, non-isotropic Measurements vbmd, microarchitecture vbmd, macroarchitecture, derived microarchitecture measurements Sites Distal radius, distal tibia Central and peripheral sites Radiation dose ~ 3 μsv ~3-8 msv FE analysis Direct voxel conversion approach Continuum level models Main High-resolution, low radiation Central sites, widely available advantages dose Main disadvantages Only peripheral sites, small typical measurement volume Radiation exposure, no direct microstructure Applying segmentation algorithms to the high-resolution images acquired by HR-QCT, makes it possible to distinguish the trabecular and cortical regions and perform a detailed analysis on each region (Boutroy et al., 2008; Buie et al., 2007; MacNeil & Boyd, 2007b). Scans take 2.8 minutes to acquire a 9.02 mm section (axial length) in the only currently available commercial system (XtremeCT, Scanco Medical, Brüttisellen, Switzerland). Radiation exposure during an HR-pQCT scan is several orders of magnitude lower than whole body CT at an approximate 3 μsv effective patient dose per scan. Automatic algorithms can be used to segment the cortical and trabecular regions with good accuracy and will be discussed in further detail (Section 2.3.1). After this segmentation, various material and structural properties can be measured and are described in section

43 Figure 2.11: HR-pQCT distal radius image with the cortical region segmented and shown in transparent gray. The cortical pores are highlighted in blue Artefacts and Limitations of CT CT artefacts can compromise image quality and thus the ability to accurately measure density or architectural characteristics. Beam hardening is one such artefact, occurring when lower energy photons are attenuated more than higher energy photons. This effect can be minimized by passing the beam through various materials as filters that absorb the low-energy photons, only allowing the high-energy photos to be passed through the patient. This process is known as pre-hardening the beam. Common filtration materials for CT may consist of layers of copper, aluminum, and carbon (Bushberg et al., 2002). Another common image artefact occurs when a voxel overlaps two distinct materials and the resulting measured attenuation reflects a proportion of those two materials (Figure 2.12). This is known as partial volume effects and may cause significant errors if the structure to be resolved is smaller than the resolution of the scanner. Finally, another common issue in CT scans is motion artefacts, particularly for HR-pQCT scans (Figure 2.12). Scans can be graded for motion either manually or automatically (Pauchard et al., 2012), and subjects should be 25

44 rescanned if the amount of motion is unacceptable. Decreasing scan time or adjusting the position of the subject to maximize his or her comfort are ways to avoid motion artefacts. Figure 2.12: CT image artefacts. The left image depicts partial volume effects when trying to represent the red circle. The image pixels that overlap the border of the object are neither dark gray nor white but a proportion of both. The right image shows a HR-pQCT scan with motion artefacts seen distinctly as streaks and breaks in the cortex. There are also some limitations of CT imaging to consider. The first is that compared to DXA scanners, CT scanners have a much higher cost and limited accessibility. Also, the radiation exposure to subjects in x-ray based modalities is an important limitation. For example, an abdomen CT scan can have an effective dose of about 8 msv, which is equivalent to about 2.7 years of normal natural background radiation (Mettler Jr et al., 2008). Reducing the beam energy in the scanner settings can minimize these doses, but at the cost of increased image noise (Dowsett et al., 2006). Magnetic resonance (MR) imaging is another modality that is beginning to be used to assess bone structure (Goldenstein et al., 2010; Krug et al., 2008) without the 26

45 exposure to ionizing radiation. However, for the application to bone quality and strength assessment, MR imaging has lower resolution than HR-pQCT, which produces more partial volume artefacts and MR imaging is not able to provide any density information. 2.3 Quantitative Image Evaluation Once medical images are acquired, image processing and evaluation are required to quantify and assess the structure and density Image Segmentation A key component in obtaining accurate measurements of BMD and architectural characteristics is accurately defining the desired regions of interest (ROI). This includes both the virtual extraction of the bone of interest, whether it is the distal radius or proximal femur, and segmentation of the cortical and trabecular regions. For HR-pQCT images, various methods have been proposed for the segmentation of cortical and trabecular bone, and can be classified as either fully automatic or semi-automatic. Many semi-automatic methods involve the operator drawing contour lines around the ROI and algorithms that fit contour lines to gradient boundaries in images (Kass et al., 1988; Laib et al., 1998). Interpolation can then be used to determine the contour lines for other slices, thus eliminating the need for all slices to be manually contoured. The major limitation of these methods is that they require significant operator interaction, which is both time consuming and prone to error. An automatic method proposed by Dufresne uses a Euclidean distance map where the distance to the nearest surface is stored for each voxel (Dufresne, 1998). Using surface normals 27

46 and the distance map, the center of the cortical bone can be found and dilated to produce a mask for the cortical region (Dufresne, 1998). A disadvantage to this method is that it can be affected by holes in the cortical structure such as cortical porosity. The dual-threshold segmentation algorithm is a fully automated method that uses a series of dilations and erosions to obtain a mask for the cortical and trabecular regions of the bone (Figure 2.13) (Buie et al., 2007). Due to the assumption that the trabecular region is always surrounded by cortical bone, it is able to detect very thin sections of the cortex. Further details are provided in section Figure 2.13: Result of an automatic segmentation by the dual-threshold method of the cortical and trabecular regions in a distal radius HR-pQCT image. The cortex is shown in blue and the segmented trabecular region is shown in gray. Due to the lower resolution of QCT images, and thus larger effects of partial volume errors, segmentation of the region of interest and cortical and trabecular regions is even more challenging. Various methods have been proposed including threshold and region growing 28

47 algorithms (Kang et al., 2003; Mahaisavariya et al., 2002). Many of these methods are very time consuming and do not distinguish between the cortical and trabecular regions. Recently, two methods have been developed to extract the bone of interest as well as accurately segment the cortical and trabecular regions. The first uses an active contouring method to find the endosteal and periosteal boundaries (Chan and Vese, 2001). The normals along the surfaces are then sampled to provide the cortical and trabecular boundary accurately to about 1/3 of the pixel size (Hangartner and Short, 2007). The second method developed by Treece and colleagues uses a threshold-based technique to extract the endosteal surface of the bone with some manual corrections, and then many independent thickness measurements are estimated across the cortex adjusting for the partial volume effects (Treece et al., 2010). This method produces an unbiased measurement of the cortex down to 0.3 mm (Treece et al., 2010) Image Calibration Another important step in obtaining accurate estimates of bone mineral density is the calibration of the Hounsfield units measured by the CT scanner. For some scanner designs, such as the HR-pQCT scanner (XtremeCT; Scanco Medical; Brüttisellen, Switzerland) that only scans peripheral sites, a daily and weekly calibration is sufficient. This calibration phantom is composed of a series of materials with known densities that is scanned separately from the subjects (Laib et al., 1998). Clinical CT scanners typically require within-scan calibration phantoms to allow for accurate quantification of densities because of the large variations of other materials that may exist in the field of view such as soft tissue. The values measured by the CT scanner in the phantoms can then be plotted against the known density values in the phantom to calibrate the values in the remainder of the image as shown in Figure Based on these 29

48 calibrations, studies have shown very high correlations between scanners at different sites and between different manufacturers (Ito et al., 2010). Recently phantom-less CT has been used for vbmd measurements, and while promising, is still inferior to phantom-based vbmd measurements. This is because the densities of the materials used for calibration in the image are only approximately known unlike the actual known densities in the calibration phantoms (Habashy et al., 2011; Mueller et al., 2011). Figure 2.14: Sample hip CT scan with calibration phantom underneath the subject (left). Enlarged portion shows the phantom used to calibrate the density measurements. On the right is an example plot to determine the linearity of the scan as well as the values to calibrate the image HR-pQCT Morphometric Indices In addition to vbmd, which can be calculated for the whole bone as well as the cortical and trabecular regions, bone morphometry can be used to assess the microstructural properties of 30

49 bone. Prior to micro-ct imaging, most morphometric indices were calculated from twodimensional histology slices. Now, if the resolution is high enough to resolve the structure being imaged, indices can be measured directly from the three-dimensional images (Hildebrand and Rüegsegger, 1997). However, since the resolution of HR-pQCT is relatively close to the size of the trabeculae, trabecular measurements are generally derived from values less affected by resolution (Laib et al., 1998). Trabecular bone volume fraction (BV/TV) is derived from trabecular vbmd where 1200 mg HA/cm 3 is used as the threshold for fully mineralized bone (Laib et al., 1998). Trabecular number (Tb.N) is the average number of trabeculae per unit length and is directly calculated as the inverse of the distance between the mid-axes of the trabecular elements. Based on Tb.N and BV/TV, trabecular thickness (Tb.Th) and trabecular separation (Tb.Sp) can be derived using the formulas shown in Table 2.2 (Boutroy et al., 2005). Since cortical thickness (Ct.Th) can be well resolved with HR-pQCT, it can be measured directly; however, depending on the segmentation, derived Ct.Th measurements are also common. The direct measurement is based on a distance transform of cortical region. The derived measurement of Ct.Th is calculated as the volume of the cortical bone divided by outer bone surface (Laib et al., 1998). As will be further explained in Chapter 3 of this thesis, cortical porosity can be measured similarly to BV/TV except that the fraction of void space, rather than bone, to the total cortical region is quantified. Both trabecular and cortical HR-pQCT measurements are highly correlated with measurements performed on cadavers by micro-computed tomography (micro-ct) measurements (Liu et al., 2010b; MacNeil and Boyd, 2007a). In addition, measurements have a 31

50 precision of <1% and <4.5% for density and morphological parameters, respectively (MacNeil and Boyd, 2008b). A summary of these measurements can be found in Table 2.2. Table 2.2: Morphological indices used to quantify bone structure with HR-pQCT. Indices Abbreviation Units HR-pQCT calculation Reference Bone volume fraction BV/TV Assumes bone density is >1200 mg HA/cm 3 and calculates (Laib et al., 1998) fraction of bone volume to total volume in trabecular region Trabecular Number Tb.N 1/mm Direct: Inverse of the mean spacing of the mid-axes (Boutroy et al., 2005) Trabecular Thickness Tb.Th mm Derived: Tb.Th = (BV/TV) /Tb.N (Boutroy et al., 2005) Trabecular Tb.Sp mm Derived: Tb.Th = (Boutroy et Separation Cortical Thickness Cortical Porosity (1 - BV/TV) /Tb.N Ct.Th mm Derived: Ct.Th = Cortical volume / outer surface area Direct: Distance transform Ct.Po % Assumes bone density is >1200 mg HA/cm 3 and calculates fraction of void volume to total volume in cortical region Total Area Tt.Ar mm 2 Average cross sectional area for whole bone Trabecular Area Tb.Ar mm 2 Average cross sectional area for trabecular bone Cortical Area Ct.Ar mm 2 Average cross sectional area for cortical bone al., 2005) (Boutroy et al., 2005; Nishiyama et al., 2010b) (Nishiyama et al., 2010b) (Boutroy et al., 2005; Nishiyama et al., 2010b) (Nishiyama et al., 2010b) (Nishiyama et al., 2010b) 32

51 2.3.4 CT Measurements and Fracture Risk The volumetric density measurements and morphologic indices described above provide a detailed overview of an individual s bone quality. Once quantified, these measurements from both QCT and HR-pQCT can be related to fracture outcomes. In a large prospective study, bone structural parameters measured by QCT were shown to be independently associated with fracture of the proximal femur (Black et al., 2008). However, the prediction of fracture was not improved when QCT indices were added to measurements of abmd by DXA (Black et al., 2008). Bousson and colleagues found that QCT structural parameters were not significantly better at predicting hip fracture compared to abmd despite the additional structural and volumetric density information provided (Bousson et al., 2011). These studies suggest that the individual parameters acquired from QCT may not be fully capable of describing the structural complexity of the proximal femur. Recently, studies have also began to examine the association between HR-pQCT measurements and fracture outcomes (Boutroy et al., 2008; Melton et al., 2007; Vico et al., 2008; Vilayphiou et al., 2010). Vico and colleagues found that BV/TV, Tb.N, Ct.Th, and Ct.BMD were significantly different between those with a hip fracture and controls (Vico et al., 2008). Additionally, cortical parameters (thickness, density, and area) were also significantly different between those with wrist fracture compared to those with hip fracture, indicating the importance of the cortical region (Vico et al., 2008). It is also important to note that multiple studies have found that bone structure contributes to fracture risk independently of abmd (Melton et al., 2007; Sornay-Rendu et al., 2007). These associations have been found for individual parameters as well as combination of parameters produced by principle component analysis (Boutroy et al., 2008; Vilayphiou et al., 33

52 2010). Two studies have shown that the first principal component accounting for the majority of the variance in the data was represented mainly by cortical bone thickness and density (Boutroy et al., 2008; Vilayphiou et al., 2010). Both of these main components were also significantly associated with wrist fracture (Boutroy et al., 2008) and fractures at all sites (Vilayphiou et al., 2010). Trabecular indices were generally accounted for in the second principal component, but did not result in significant odds ratios for fracture (Boutroy et al., 2008; Vilayphiou et al., 2010). While these measurements are promising on their own, more direct estimates of bone strength may further improve fracture prediction. 2.4 Bone Mechanics Structural and microarchitectural indices have been shown to be useful surrogates for bone strength and in some cases are associated with fracture outcomes. Since fractures occur when the applied load exceeds the mechanical strength (Keaveny and Bouxsein, 2008), it is important to also provide estimates of bone strength to assess fracture risk Basic Mechanics Force represents an action on an object and is composed of a magnitude and a direction. In order to characterize a material, force can be plotted against the material s displacement under a given load, which depends on the structure (i.e. size and shape) of the material (Figure 2.15). Stress is a force acting on a specific area with a unit of measure of either pascals (Pa) or newtons per square meter (N/m 2 ) while strain is a dimensionless measure of the deformation per unit length due to a force. Unlike the load-displacement plots, a plot of stress versus strain does not depend on the structure of the object in question (Figure 2.15). 34

53 Various properties can be used to characterize a material and its resistance to fracture. The slope of the load-deformation plot in the linear region represents the stiffness of the material (Beer et al., 2004). The equivalent region from the stress-strain plot is also known as the Young s, or elastic modulus. Other intrinsic properties include the point where the line transitions from the elastic to plastic region known as the yield stress, and the highest point on the curve before fracture is the ultimate or maximum stress. The toughness of a material is the area under the entire curve and represents the amount of energy the material can absorb before complete failure occurs while the resilience of a material is the area under the curve up to the yield stress and represents that energy that can be absorbed before yielding occurs (Beer et al., 2004). Figure 2.15: Load-deformation plot (left) and stress-strain plot (right) depicting the various properties that can be used to characterize a material. 35

54 Mechanical testing can be used to experimentally determine mechanical properties of an object. These tests can be either force or displacement controlled where one of the variables is applied and the other is measured to produce a force-displacement plot. Many considerations such as temperature, hydration, and storage must be addressed when performing mechanical testing in order to not alter the mechanical properties (Langton and Njeh, 2004). For example, compared to a hydrated bone, a dry bone may have increased Young s modulus but decreased toughness (Cowin, 2001). The loading configurations lead to specific fracture patterns. For example, compression causes oblique fractures, tension causes transverse fractures, and torsion produces spiral fractures. These loads can be complex in nature and may represent a combination of these conditions. Bending, which is commonly observed in a sideways fall on the hip, is a combination of both tension and compression produces a transverse fracture with a small fragment on the concave side (Einhorn, 1992) Bone Mechanical Properties Bone fragility due to decreased bone strength results due to a decrease in both the mass of bone as well as the deterioration of the structure. A careful balance of stiffness and flexibility are required to withstand loads and deform for energy absorption. In biological materials such as bone, the strength depends on both the organic and inorganic components. The organic collagen matrix provides flexibility to the bone and the inorganic mineral deposits contribute to the stiffness. An imbalance in these components may lead to a bone that is very brittle (e.g. osteopetrosis) or under mineralized (e.g. osteomalacia) both of which increase the risk of fracture (Figure 2.16) (Cowin, 2001). 36

55 Figure 2.16: Example load-deformation plots for different diseases showing varying stiffness as well as proportions of elastic and plastic regions. Adapted with permission from (Sato et al., 1999). Copyright American Chemical Society. On a larger scale, the amount and structure of the cortical and trabecular bone also determines the mechanical strength (Liu et al., 2010b; MacNeil & Boyd, 2007b; Sornay-Rendu et al., 2007). The number, thickness, and arrangement of the trabeculae are the main determinants of the mechanical strength of the trabecular region (Seeman, 1999) while the mechanical strength of the cortex is mainly determined by the cross-sectional area, thickness, and porosity (MacNeil & Boyd, 2007b; Spadaro et al., 1994). Various types of mechanical tests can be performed to measure the mechanical properties of bone including compression, torsion, and bending and often involve a specific sample of material (i.e. a cube or cylinder of bone). In addition, more complex, site-specific tests can be 37

56 performed in which specific boundary conditions can be applied to simulate in vivo forces or forces occurring during specific loading situations (i.e. a fall) (Langton & Njeh, 2004) Site-Specific Bone Mechanical Testing The distal radius is one of the most common sites of fracture and occurs due to a fall on the outstretched arm (Chung and Spilson, 2001). Consequently, mechanical tests have been designed to simulate these forces to measure the bone strength (Eckstein et al., 2004). Some studies have used full cadaver arms or whole bones to perform these tests (Muller et al., 2003; Pistoia et al., 2002) while many others have simplified this test to compression of a slab from the distal portion of the radius (Eckstein et al., 2002; Lochmüller et al., 2002; MacNeil and Boyd, 2008a). For the proximal femur there are two main loading conditions that simulate the forces applied to the bone (Figure 2.17). The first is a standing configuration where the load to the femur is applied to the top of the femoral head as would be encountered when standing or walking (Keyak, 2000). In this loading configuration, failure loads are of the order of approximately 5000 N (Keyak, 2001). Since over 90% of fractures occur during falls (Grisso et al., 1991; Nevitt et al., 1989), it is important to determine the strength of a bone in a falling configuration. The estimated bone failure load for this configuration are much lower compared to the standing configuration, around 2000 N (Keyak, 2001). 38

57 Figure 2.17: Falling (left) and standing (right) loading configurations for the mechanical testing of the proximal femur. Force arrows are indicated in red. Based on these tests, force-displacement curves can be generated to measure the stiffness and ultimate load of the specimens. If the original size and cross-sectional area is known, these load-deformation curves can be used to generate a stress-strain plot (Beer et al., 2004). The loading rate of the mechanical tests can have an effect on the outcome measurements because of the viscoelastic properties of bone. It has been shown that as the strain rate increases so does the measured Young s modulus (McElhaney, 1966). Strain rates of 0.01 to 0.08/s represent the range of strains occurring in vivo (Rubin and Lanyon, 1982). Although these methods provide direct measurement of the mechanical properties, they are destructive tests and often prone to error (Odgaard and Linde, 1991). 39

58 2.4.4 Finite Element Analysis While mechanical testing is the gold-standard approach for measuring the strength of bones, it is not practical to apply to patients due to its destructive nature. Alternatively, nondestructive, in vivo approaches such as finite element (FE) analysis can be used to estimate bone strength. FE analysis is a modeling technique where a mechanical problem is solved numerically. A continuum model is discretized into individual elements and specific boundary conditions can be applied. Like mechanical testing, the model is either solved for force or displacement. These models can then be validated by experimental mechanical tests to determine their accuracy. Linear approximations can be used for FE analysis and have the important advantages of being easier to compute and much less computationally intensive. These models represent the elastic region of the material and thus only provide information such as the stiffness of the material. In some applications it has been shown that stiffness from linear analysis can be used indirectly to estimate ultimate strength (MacNeil & Boyd, 2008a). In order to estimate failure or to achieve higher accuracy, nonlinear FE analysis can be used. This can be either geometric nonlinearities caused by large displacements, strains, or rotations, or material nonlinearities where the force-displacement plot is not linear. The major drawback of nonlinear FE analysis is that it can be extremely computationally expensive. Tradeoffs such as reducing the image resolution, and consequently model resolution, can be used as compromises for nonlinear analysis (Koivumäki et al., 2012) Image-Based Finite Element Analysis Traditional FE analysis techniques require a user to generate a mesh of the object in question that is then used as the input for the FE analysis. Another technique uses three- 40

59 dimensional data, such as medical images, to directly generate the mesh from the image (Müller & Rüegsegger, 1995; Van Rietbergen et al., 1995). Each image voxel can be directly converted to a finite element in the mesh. Since every element has the same stiffness matrix, only one copy of the matrix needs to be stored resulting in a memory savings on the order of 99.9% (Smith and Griffiths, 1998). Combined with matrix preconditioning and parallelizing the software, significant time savings can be achieved (Van Rietbergen et al., 2003). Once these meshes are solved, subject-specific estimates of bone strength can be determined. Both QCT (Cody et al., 1999; Keaveny et al., 2010; Keyak et al., 1998) and HR-pQCT images (Boutroy et al., 2008; Burghardt et al., 2010a; MacNeil & Boyd, 2008a) have been used to generate these FE models (Figure 2.18). In the case of QCT images, since the resolution is not able to resolve specific structures, the image voxels overlap various densities (i.e. partial volume effects). Thus, the FE models are based on average CT attenuation values and converted into corresponding tissue moduli based on an appropriate density to modulus conversion (Helgason et al., 2008). There have been various density-to-modulus relationships proposed based on experimental tests (Table 2.3) (Helgason et al., 2008). The most appropriate relationship often depends on the specific application (i.e. cortical or trabecular bone, site of bone, or test conditions). 41

60 Table 2.3: Various density (ρ) to modulus (E) relationships reported in previous studies. R 2 is the determination coefficient. A more comprehensive list can be can be found in (Helgason et al., 2008). Study Site of Bone Type of Bone Density to Modulus R 2 (Keller, 1994) Proximal femur Cortical and E = 10.5ρ trabecular (Keyak et al., 1994) Proximal tibia Trabecular E = 33.9ρ (Keaveny et al., 1997) Lumbar spine Trabecular E = 1.540ρ (Morgan et al., 2003) Vertebrae Trabecular E = 4.730ρ (Morgan et al., 2003) Femoral neck Trabecular E = 6.850ρ Using HR-pQCT images, the resolution is high enough that the structure can be represented directly by the elements in the model. Images can be binarized and all bone elements can be assigned a homogenous elastic modulus. In specific applications where mineralization changes may be expected, models can incorporate tissue inhomogeneities based on the x-ray attenuation values (Bourne and van Der Meulen, 2004). Due to the extremely large size of these models, custom FE solvers are required (MacNeil & Boyd, 2008a; Van Rietbergen et al., 1995). 42

61 Figure 2.18: An HR-pQCT tibia scan with 1% uniaxial compression applied (far left). Other images showing various views of the FE model coloured by Von Mises stress (MPa) Bone Strength Estimates and Fracture The estimates of bone strength from the previously described FE models can be and used towards the ultimate goal of determining fracture risk in a specific patient. While the CT measurements described in section have associations with fracture, it has also been reported that some bone treatments may increase average BMD while others increase bone strength (Keaveny et al., 2007). The latter may be more important to consider when determining fracture risk. Using QCT images and FE analysis, investigators have examined the association between bone strength estimates and hip fracture (Amin et al., 2011; Keyak et al., 2011; Orwoll et al., 2009). Some studies have shown that FE models are better able to predict femoral strength than QCT-based BMD and DXA-based abmd (Cody et al., 1999; Crawford et al., 2003). In addition, 43

62 estimates of bone strength were found to be strongly associated with fracture in both men (Amin et al., 2011; Keyak et al., 2011; Orwoll et al., 2009) and women (Amin et al., 2011; Keyak et al., 2011). Orwell and colleagues (2009) reported that bone strength was not an independent predictor of fracture after accounting for abmd but the load-to-strength ratio (which incorporates fall force and estimated strength) remained a significant predictor. Amin et al. (2011) similarly found that FE estimates of bone strength were not better than abmd or vbmd at discriminating fracture (Amin et al., 2011). One study has shown that trochanteric fractures are more sensitive to loading direction than cervical neck fractures (Wakao et al., 2009), which suggests that estimating bone strength at various loading configurations may improve fracture prediction. Based on FE models generated from HR-pQCT scans, associations with wrist fractures (Boutroy et al., 2008) and fractures at other sites (Vilayphiou et al., 2010) have been studied. Boutroy and colleagues (2008) reduced the FE and microarchitecture parameters into principal components that account for the majority of the variation in the data. The first principal component, which had the highest odds ratio with fracture (2.49, CI: ), included FE estimated stiffness. Vilayphiou et al. (2010) extended this study to several types of fragility fractures and similarly found that FE stiffness had the highest weight in the first principal component. In a separate cohort of women with wrist fractures, FE bone strength alone was found to have a significant association (OR: 3.8; CI: ) with fracture (Melton et al., 2007). Importantly, finite element estimates of bone strength using HR-pQCT have also been shown to provide addition information beyond that of abmd or bone microarchitecture alone (Boutroy et al., 2008). 44

63 2.5 Machine Learning Classification As discussed above, most previous studies have focused on finding the association between microarchitecture measurements and bone strength estimates with fracture. The next step is to use classification models that can predict the individuals that are at risk of fracture Association Models Association models provide insight into the relationship that risk factors or diagnostic tests have with a disease. This is useful for indicating potential interventions to prevent or treat a disease; however, these studies provide insights at the population level rather than individual patient level (Feng, 2010). Odds ratios are typically the main outcome for association studies to indicate the association between a measured parameter and binary outcome (i.e. fracture). Other outcomes often include hazard ratios, when time is the outcome, or correlation coefficients for a continuous outcome. One important limitation of association models is that they do not provide information regarding the sensitivity or specificity of a method, which is often an important consideration for clinical tests (Pepe et al., 2004) Classification Models Classification models provide insight into the subject-level performance of a measure to predict an outcome. Often good associations are needed, but are not sufficient on their own, for a good classification (Feng, 2010; Pepe et al., 2004). These models can be the basis for clinical decision making for individual patients. The outcomes of classification models reflect the performance related to decision making. The sensitivity of a test indicates the probability of a model detecting an abnormal event 45

64 (i.e. a fracture) based on the number of true positives (TP) and false negatives (FN). The specificity indicates the probability that a test correctly identifies a negative case (i.e. nonfractured) based on the number of true negatives (TN) and false positives (FP). The accuracy of the classification is also normally reported but on its own is not always a good indicator, particularly in the case where a classifier results in all positive or all negative results. Equation 2.1: Performance measures for classification models based on true positives (TP), true negatives (TN), false negatives, and false positives (FP). Receiver operating characteristic (ROC) curves provide a performance measure of the false positive rate (1-specificity) versus the true positive rate (sensitivity) (Zweig and Campbell, 1993). ROC curves are presented as a two-dimensional plot to depict how the classifier varies as different thresholds are used. The overall performance is often summarized as the area under the curve (AUC). Based on these outcomes, decisions for individuals can be made according to the specific application (Zweig & Campbell, 1993). For accuracy and ROC-AUC measurements, cross-validation schemes can be used to ensure the model is not over fit to the specific dataset. For example, an N-fold cross-validation would divide the dataset into N equal groups, train the model on N-1 of these groups, and test on the remaining group. This is repeated so that every group is used as the testing group (Kohavi, 1995). 46

65 2.5.3 Machine Learning Methods Several machine-learning methods exist for the classification of datasets each with their own advantages and disadvantages. Some of the most prominent include artificial neural networks, gradient boosting machines, and support vector machines (SVM s). Artificial neural networks are statistical models that are designed to mimic the connections in biological neural networks (Dayhoff and DeLeo, 2001). They can be trained to model complex systems with multiple layers of connected neurons. Artificial neural networks have been used for the prediction of subjects with osteoporosis based on demographic and anthropometric characteristics with very good performance (AUC = 0.82) (Chiu et al., 2006). Boosting is a method in which multiple prediction rules are combined to improve the overall prediction (Schapire, 2003). In the case of gradient boosting machines, multiple classifications by decision trees are combined into a single model (Friedman, 2002). In a very recent study, Atkinson and colleagues demonstrated excellent performance when classifying subjects with fracture based on QCT, HR-pQCT, and DXA measurements using gradient boosting machines (Atkinson et al., 2012). Support vector machines attempt to find the maximum margin between groups within a dataset (Figure 2.19) (Cortes and Vapnik, 1995). The data is mapped to a higher dimensional space to allow for nonlinear boundaries resulting in improved accuracy and efficiency (Burges, 1998; Cortes & Vapnik, 1995). Recently, SVM s have been used to discriminate subjects with fracture based on texture measurements from x-ray images (Lee et al., 2008). 47

66 Figure 2.19: Example of a simple two-dimensional support vector machine classification. The dashed line shows the optimal hyperplane margin to separate the two classes (squares and circles) and the support vectors are shown in red. Adapted from (Cortes & Vapnik, 1995) with permission. Compared to neural networks, SVM s have the important advantage of finding a global unique solution rather than local minima (Burges, 1998). In many cases gradient boosting machines have shown to have very high classification accuracy but compared to SVM s they require significant tuning of the parameters (i.e. number of trees and shrinking parameters), which can lead to a very high computational cost (Johnson and Zhang, 2011). 2.6 Knowledge Gaps and Summary This chapter has provided relevant background information and has highlighted potential research opportunities. As medical imaging technologies advance, improvements continue to be made to the assessment of bone quality and strength with the ultimate goal of improving fracture 48

67 prediction. This thesis focuses on the development and validation of methods to extract and interpret the material and structural properties that can now be measured. Using HR-pQCT, much more work has focused on analyzing trabecular bone (Boutroy et al., 2008; Hildebrand and Rüegsigger, 1997; Liu et al., 2010a; MacNeil & Boyd, 2007a) than cortical bone. Since cortical bone makes up the majority of bone mass in the body and is a significant contributor to bone strength (Cowin, 2001), it is equally important to study the structural characteristics of cortical bone in high fracture risk populations. One of the structural properties that has not yet been studied in vivo is cortical porosity even though several in vitro studies have indicated that it is an important predictor of bone strength (McCalden et al., 1993; Wachter et al., 2001a; Yeni et al., 1997). In addition, porosity increases with age (Bousson et al., 2001; Stein et al., 1999a) and is high during puberty (Parfitt, 1994b); both of which are times of high fracture incidence. Until recently, it has not been possible to quantify cortical porosity in vivo because of the high-resolution needed to resolve the pores as well as the difficulties associated with segmenting the cortical and trabecular regions. Quantitative CT also shows promise to improve the assessment of bone strength and fracture risk. The main strength of QCT is its ability to measure structure and geometry at central sites within the body. Evidence of the validity of FE analysis to estimate bone strength has been presented (Bessho et al., 2007; Dragomir-Daescu et al., 2011; Keyak et al., 1998; Koivumäki et al., 2012; Viceconti et al., 2004); however, clinical relevance has been limited due to the time and computational power required to solve the models. This restriction is even greater when trying to simulate multiple loading configurations for each subject: a technique that may provide greater insights into fracture risk since there are many different ways in which a person may fall. 49

68 While association studies have been completed examining the relationship between fracture and bone strength estimates based on HR-pQCT (Boutroy et al., 2008; Vico et al., 2008; Vilayphiou et al., 2010) and QCT (Amin et al., 2011; Keyak et al., 2011; Orwoll et al., 2009), there is still a need for classification studies to provide insight into the ability of these methods to distinguish fractures in individual subjects. The following chapters focus on filling these knowledge gaps and addressing the objectives presented in Chapter 1 in a manuscript-based format. 50

69 CHAPTER THREE Cortical Porosity in Pre- and Postmenopausal Women 51

70 CHAPTER THREE: POSTMENOPAUSAL WOMEN WITH OSTEOPENIA HAVE HIGHER CORTICAL POROSITY AND THINNER CORTICES AT THE DISTAL RADIUS THAN WOMEN WITH NORMAL abmd: AN IN VIVO HR-pQCT STUDY This chapter is reprinted with permission from a manuscript published in the Journal of Bone and Mineral Research: Nishiyama KK, Macdonald HM, Buie HR, Hanley DA, Boyd SK. Postmenopausal women with osteopenia have higher cortical porosity and thinner cortices at the distal radius and tibia than women with normal abmd: an in vivo HR-pQCT study. J Bone Miner Res 2010; 25: Author Roles: KKN contributed to the conception, design, data analysis, interpretation, and drafting of the manuscript. HRB contributed to technical analysis. HMM, DAH, SKB contributed to the conception, design, and interpretation. All authors critically revised the manuscript. Abstract Increases in cortical porosity (Ct.Po) and decreases in cortical thickness (Ct.Th) are associated with increased bone fragility. The purpose of this study was to validate an autosegmentation method for high-resolution peripheral quantitative computed tomography (HRpQCT) scans to measure Ct.Po and Ct.Th, and use it to compare Ct.Po and Ct.Th between preand postmenopausal women with normal, osteopenic and osteoporotic areal bone mineral density. The Ct.Po and Ct.Th measurements were validated using cadaver forearms (n=10) and micro-computed tomography as the gold standard. The analysis was applied to distal radius and tibia HR-pQCT scans from a subset of participants from the Calgary, AB cohort of the Canadian 52

71 Multicentre Osteoporosis Study (n=280, yrs). Analysis of covariance compared Ct.Po and Ct.Th outcomes between 63 normal premenopausal (DXA femoral neck T-score > -1), 87 normal postmenopausal, 121 osteopenic postmenopausal, and 9 osteoporotic postmenopausal women. Linear regression analysis and Bland-Altman plots were used to assess the agreement between the HR-pQCT and µct measurements, resulting in R 2 values of 0.80 for Ct.Po, and 0.98 for Ct.Th. At both sites Ct.Po was higher in postmenopausal (all groups) than premenopausal women (3.2 to 12.9%, p<0.001). Ct.Th was not significantly different between normal premenopausal and postmenopausal women at either site; however, both osteopenic and osteoporotic women had thinner (-12.8 to -30.3%, p<0.01), more porous (2.1 to 8.1%, p<0.001) cortices than normal postmenopausal women. Our method offers promise as a valuable tool to measure Ct.Po and Ct.Th in vivo and investigate associations between cortical bone structure, age and disease status. 3.1 Introduction Cortical bone contributes significantly to whole bone strength (Augat and Schorlemmer, 2006; Holzer et al., 2009; MacNeil & Boyd, 2007b; Spadaro et al., 1994); however, the contribution of the cortex to bone strength is often overlooked (Augat & Schorlemmer, 2006). The strength of cortical bone is determined by both material properties and the microarchitecture (Müller, 2002). For example, cortical porosity which occurs as the result of bone resorption on the endocortical surface and on the surfaces of Haversian canals (Epker and Frost, 1965; Foldes et al., 1991; Parfitt, 1994a) is a significant predictor of bone s mechanical strength (Currey, 1988; McCalden et al., 1993; Schaffler and Burr, 1988; Wachter et al., 2001b; Yeni et al., 1997). Findings from in vitro histology and micro-computed tomography (µct) studies indicate that 53

72 cortical porosity increases with age (Bousson et al., 2000; Busse et al., 2010; Cooper et al., 2007a; McCalden et al., 1993; Stein et al., 1999b; Ural and Vashishth, 2007) and this age-related increase was shown to account for 76% of the loss in cortical bone strength at the clinically relevant proximal femur (McCalden et al., 1993). Further, cortical porosity at the femoral neck was found to be higher in individuals who suffered a hip (Barth et al., 1992; Bell et al., 2000; Squillante and Williams, 1993) fracture compared with age-matched subjects who were fracturefree. More recently, vertebral fractures in men with idiopathic osteoporosis were found to be associated with increased cortical porosity as measured with histomorphometry at the iliac crest (Ostertag et al., 2009). In addition to cortical porosity, structural and geometric properties of the cortex such as cortical thickness are important determinants of bone strength (Augat & Schorlemmer, 2006). For example, in biomechanical studies, cortical thickness (or width) was found to be a significant predictor of fracture load at the distal radius (Augat et al., 1996). Furthermore, thinning of the cortex occurs with ageing (Khosla et al., 2006; Mayhew et al., 2005) and disease (Boutroy et al., 2005) and has been shown to result in bone fragility and increased fracture risk (Mayhew et al., 2005). Despite the well-documented contribution of cortical bone to whole bone strength, in vivo assessments of cortical bone microstructure including cortical porosity have been limited by the high image resolution required and difficulties in segmenting the cortex from the trabecular region. With the advent of high-resolution pqct (HR-pQCT), trabecular and cortical microarchitecture of the distal radius and tibia can be safely measured at 82 µm nominal isotropic resolution (Liu et al., 2010c; MacNeil & Boyd, 2007a) and automated segmentation algorithms can be employed to separate the cortical and trabecular bone compartments (Buie et 54

73 al., 2007). HR-pQCT has been used to evaluate cortical bone properties in the adolescent (Kirmani et al., 2009) and adult (Boutroy et al., 2005) skeleton; however, current segmentation software cannot be used to accurately measure porosity. In addition, current methods used to define porosity (Kirmani et al., 2009) have not been validated. Therefore, the purpose of this study was to validate HR-pQCT measures of cortical porosity and thickness using an automated segmentation algorithm, and apply this analysis to a population-based sample to characterize cortical bone in normal pre- and postmenopausal women, osteopenic postmenopausal and osteoporotic postmenopausal women. 3.2 Materials and Methods In Vitro Validation Specimens Human cadaver forearms (5 men, 5 women; years) were obtained from the Gross Anatomy Lab at the University of Calgary and were kept frozen until 48 hours before scanning. Medical history was unavailable and thus the presence of previous bone diseases could not be determined. All procedures were approved by the Office of Medical Bioethics at the University of Calgary HR-pQCT and µct Scanning The intact forearms were first scanned with HR-pQCT (XtremeCT, Scanco Medical, Brüttisellen, Switzerland) at four adjacent locations using the standard manufacturer s in vivo parameters (60 kvp, 1000 µa, 100 ms integration time) as described previously (MacNeil & Boyd, 2007a). The second most distal section corresponded to the standard in vivo measurement 55

74 site at 9.02mm from the most proximal location of the sub-chondral plate Figure 3.1 (MacNeil & Boyd, 2007a). Each section encompassed 110 slices and corresponded to 9.02 mm measured with an 82 µm voxel size. Figure 3.1: Scout views representing the reference line (solid line) and region scanned (dotted lines) at the distal radius (A.) and distal tibia (B.). The radii were scanned using µct (vivact 40, Scanco Medical, Brüttisellen, Switzerland), which is a validated and reliable tool for in vitro quantification of cortical porosity in human bones (Cooper et al., 2004b; Wachter et al., 2001b), and was used as our gold-standard. The radii were dissected free of soft tissue and scanned in a saline solution using in vitro scanning parameters (55kVp, 109µA, 300ms integration time) and a 19 µm voxel size. A 38 mm region was scanned and overlapped the entire region scanned by HR-pQCT. 56

75 Image Registration and Processing Three-dimensional image registration was used to align the images to ensure that the same regions were analyzed by both HR-pQCT and µct. Multi-resolution rigid registration using mattes mutual information and a linear interpolator (Mattes et al., 2001) were used as described in previous studies (Boyd et al., 2006). The HR-pQCT images were filtered (Laplace-Hamming filter) and binarized using the standard manufacturer s method (Laib et al., 1998). The cortical and trabecular regions were segmented using an automated segmentation method (Buie et al., 2007) implemented in Image Processing Language (IPL V5.07, Scanco Medical, Brüttisellen, Switzerland). The automated segmentation uses two threshold values and a series of morphological operations (e.g., dilation and erosion operations) to extract the endosteal and periosteal surfaces of the cortex. This is based on the assumption that the trabecular region is enclosed by a cortical shell. The standard manufacturer s method of using a Gaussian filter and threshold (filter-threshold) (Laib et al., 1998) was also applied for comparison. Due to hardware constraints, the Laplace-Hamming filter could not be used on the 19µm µct images. Therefore, they were filtered using a Gaussian filter (sigma = 1.2, support = 2) and binarized using a global threshold of 18.4% of the maximum value which is a standard analysis method for µct images. The cortical and trabecular regions were segmented by semi-automatic, hand drawn contours around both the endosteal and periosteal surfaces Cortical Bone Measurements Cortical porosity (Ct.Po, %) was calculated as the number of void voxels in each binary cortex image divided by the total number of voxels in the cortex using IPL. Cortical thickness 57

76 (Ct.Th, mm) was measured directly by removing the intracortical pores from the binary cortex image and using a distance transform (Hildebrand & Rüegsegger, 1997). In addition, Ct.Th was determined using the standard manufacturer s method, which divides the cortical area by the circumference of the cortex (Hildebrand & Rüegsegger, 1997), for comparison. The number of individual pores was counted using component labeling (IPL, Scanco Medical) and the mean pore volume was calculated as the total volume of porosity divided by the pore number. Apparent cortical density (D cort, mg HA/cm 3 ) was determined by converting attenuation values to hydroxyapatite densities using the cortical masks generated from the automatic segmentation method In Vivo Assessment Subjects The in vivo HR-pQCT scans were performed on a subset of women from the Calgary, AB, cohort of the Canadian Multicentre Osteoporosis Study (CaMos). Briefly, CaMos is a 10- year prospective population-based study of more than 9000 adult women and men from nine centres across Canada (Kreiger et al., 1999). Participants were recruited using a stratified random sampling technique that is described in detail elsewhere (Kreiger et al., 1999). At the 10-year follow-up, individuals from the Calgary CaMos cohort were invited to participate in a HR-pQCT sub-study of age-related variation in bone micro-architecture and strength. Of 621 participants (women and men), 271 (44%) women aged yrs. volunteered. To recruit more participants in the younger age strata (20-60 yrs.), three sampling strategies were employed: 1) random sampling using CaMos sampling techniques (Kreiger et al., 1999); 2) snowball recruitment whereby current volunteers were asked to provide names of 58

77 friends and family in the target age range who might be interested in participating, and 3) advertising with posters. A total of 107 women aged were recruited using these methods. Height, weight, medication use and menopause status were determined from the interviewer-administered CaMos questionnaire. The cohort included 115 premenopausal women and 263 postmenopausal women, 75 of whom had undergone a hysterectomy and/or oophorectomy. In addition, as part of the standard CaMos assessment, areal bone mineral density (abmd, g/cm 2 ) of the proximal femur was measured using dual energy X-ray absorptiometry (DXA, Hologic QDR4500, Bedford, Massachusetts). Femoral neck (FN) T-scores were generated based on CaMos normative data (Tenenhouse et al., 2000) and women were classified as osteopenic or osteoporotic according to WHO categories (World Health Organization, 1994). After excluding those women who did not have a DXA scan (n=23), women who reported longterm use of glucocorticoids (>3 months, n=38) and premenopausal women who were over 55 years (n=5) or osteopenic based on their FN T-score (n=32) the final sample included 63 premenopausal women with normal FN abmd, 87 postmenopausal women with normal abmd, 121 osteopenic postmenopausal women, and 9 osteoporotic postmenopausal women In Vivo HR-pQCT Scanning All participants were scanned at the non-dominant radius and the left tibia unless there was a previous fracture at the desired site, in which case the opposite limb was scanned. The HRpQCT scans were acquired and analyzed by one of three trained operators using the standard manufacturer s in vivo parameters and auto-segmentation method described above. Images with significant motion artefacts resulting in blurring and discontinuities in the cortical shell were excluded from the present analysis (11 radius and 4 tibia scans). Thus, the present analysis 59

78 includes 62 radius and 61 tibia images for the normal premenopausal group, 82 radius and 86 tibia images for the normal postmenopausal group, 118 radius and 120 tibia images for the osteopenic postmenopausal group, and 7 radius and 9 tibia images for the osteoporotic postmenopausal group Statistical Analysis All analyses were performed using SPSS Version 17.0 (SPSS Inc, Chicago, IL). To validate HR-pQCT measures of cortical porosity and thickness, linear regression was used to determine the R 2 coefficient between the gold-standard µct measurements and the HR-pQCT measurements. Bland-Altman plots were used to qualitatively compare the bias of the measurements to the gold standard (Bland and Altman, 1986). To compare cortical porosity and thickness across the four groups of women, an analysis of covariance (ANCOVA) was performed using age, weight and height as covariates. In the case of a significant main effect of group, post-hoc pair-wise comparisons were performed and the Bonferroni correction was applied to account for multiple comparisons. 3.3 Results In Vitro Validation Cortical porosity and thickness by HR-pQCT using the auto-segmentation method were strongly predictive of the gold standard µct measurements (R 2 = 0.80 and 0.98, respectively, p<0.01, Table 3.1 and Figure 3.2). Cortical thickness measurements by auto-segmentation of the cortical and trabecular regions were more strongly predictive of the gold standard than the threshold-filter method, which had an R 2 of 0.90 with µct (p<0.001, slope = 0.96, intercept = 60

79 0.32). However, Ct.Th measurements from both methods were highly correlated (r = 0.94). With the auto-segmentation method, pore number measured by HR-pQCT was moderately predictive of pore number by µct; however, this relationship was not statistically significant (R 2 = 0.27, p=0.13). In contrast, average pore size was strongly predictive of µct measures (R 2 = 0.66, p<0.01). Bland-Altman plots indicated an overestimation of Ct.Po (6.3%), Ct.Th (0.05 mm), and pore size (0.004 mm 3 ) and an underestimation of pore number by HR-pQCT with the autosegmentation method when compared with µct. Using the filter-threshold method, Ct.Th was underestimated by 0.45mm when compared with Ct.Th by µct (Figure 3.2). Table 3.1: Regression analysis results between µct and HR-pQCT measures, all values p<0.01. Parameter R 2 Slope Intercept Ct.Po (%) Ct.Th (Auto-segmentation) (mm) Ct.Th (Manufacturer's method) (mm) Number of pores Average volume of pores (mm 3 )

80 (Table continued on following page) 62

81 Figure 3.2: Comparison of cortical bone measurements from HR-pQCT with the gold standard µct. Regression analyses are shown on the left and Bland-Altman plots on the right for cortical porosity (Ct.Po), cortical thickness (Ct.Th), number of pores, mean volume of pores, and cortical density. Solid lines indicate 95% confidence intervals In Vivo Assessment Descriptive characteristics of the groups are provided in Table 3.2 and group comparisons of the HR-pQCT outcomes are provided in Table 3.3. Compared with premenopausal women, all postmenopausal women had significantly higher Ct.Po at both the radius (range 3.2 to 9.8%, p<0.001) and tibia (range 4.8 to 12.9%, p<0.001). Osteopenic and osteoporotic women also had higher Ct.Po than normal postmenopausal women (range 2.1 to 63

82 8.1%, p<0.001). These differences in Ct.Po were mainly due to greater mean pore volume, which was significantly higher in all postmenopausal women compared with premenopausal women (range 49.7% to 143.8%, p<0.01). Differences in pore number were observed at the distal radius when the osteopenic group was compared with the normal postmenopausal group (5.8% lower pore number, p<0.01) and at the distal tibia when the osteopenic and osteoporotic groups were compared with the normal groups (range 9.9% to 18.9% lower pore number, p<0.001). Ct.Th was significantly lower in osteopenic and osteoporotic women compared with the normal pre- and postmenopausal women (range to -31.5%, p<0.01); however, Ct.Th was not significantly different between normal premenopausal and normal postmenopausal women. Cortical density differences showed similar magnitude trends to Ct.Po and the two measures were strongly correlated (r = 0.90). Cortical density was significantly lower at both sites in osteopenic and osteoporotic women compared with pre- and postmenopausal women with normal abmd (range -7.0 to -23.7%, p<0.001). Cortical density was also significantly lower in postmenopausal women with normal abmd when compared with normal premenopausal women (radius: -5.6%, tibia: -8.9%, both p<0.001) and significantly lower in osteoporotic postmenopausal women when compared with osteopenic postmenopausal women (radius: -5.9%, tibia: -7.3%, both p<0.05). 64

83 Table 3.2: Descriptive characteristics of the subjects by group (mean ± SD). 65

84 Table 3.3: Outcome mean ± SD and percent differences between groups. Pre- refers to premenopausal and Post- refers to postmenopausal. a-c. Significant difference between groups: a. P<0.001, b. P<0.01, c. P<0.05, NS, Not significant with Bonferroni correction. 66

85 3.4 Discussion In this study, new automated methods were validated to measure Ct.Po and Ct.Th in vivo using HR-pQCT. These methods were then applied to a population-based sample to compare cortical bone porosity and thickness between pre- and postmenopausal women with normal abmd (by DXA) and postmenopausal osteopenic and osteoporotic women. Our results indicate that Ct.Po at both the distal radius and distal tibia is different between pre- and postmenopausal women with normal abmd; however, Ct.Th is not related to menopause status. In addition, osteoporosis is associated with both an increase in Ct.Po and a decrease in Ct.Th. There are methodological limitations in this study that should be noted. First, the resolution of HR-pQCT is insufficient to capture the small cortical pores as pore diameter has been shown to range in diameter from 30 µm to 400 µm (Wachter et al., 2001b). This is reflected in the poor ability of HR-pQCT to predict pore number measured by µct (Table 3.1). On the other hand, our findings show that HR-pQCT is able to accurately measure total porosity (i.e., pore volume) suggesting that larger pores detected by HR-pQCT are dominant in the total porosity measures. Second, the sample size in our validation study was small (n=10 radii) and represented mostly aged bone, which has higher average porosity than younger bone. Finally, human cadaver radii were used in the validation study and as a result there were no motion artefacts in the HR-pQCT or µct images. Subject motion is a challenging issue in the acquisition of in vivo HR-pQCT images and investigation continues to determine how cortical porosity and thickness measures are affected by motion. 67

86 3.4.1 Validation of HR-pQCT for Cortical Porosity and Thickness Our automated segmentation algorithm for determining Ct.Po from in vivo HR-pQCT scans showed good agreement with the µct gold standard, although our method appeared to slightly overestimate Ct.Po and Ct.Th. Cortical porosity measures were also overestimated in previous µct studies when compared to histology (Cooper et al., 2007b) and this is likely due to the limited resolutions (Figure 3.3). We also observed reasonable agreement between HR-pQCT and µct measures of average pore size; however, the same was not true for pore number. The lack of agreement between µct and HR-pQCT measures of pore number may be due to artefacts created during hand contouring of the µct images if small gaps are left between the contour line and endosteal surface; however, the most likely explanation, as mentioned previously, is the inability of HR-pQCT to resolve the small pores. Thus, the overestimation of average pore size by HR-pQCT may contribute to the accuracy of total Ct.Po when compared to µct. 68

87 Figure 3.3: Three-dimensional images of the same region scanned with µct (A.) and HR-pQCT (B.) visually showing the resolution differences. The voxels sizes are 19µm and 82µm, respectively. For Ct.Th, both our auto-segmentation method and the filter-threshold method are in good agreement with each other; however, our auto-segmentation method was better at predicting the gold standard. This was likely a result of better extraction of the cortex by the auto-segmentation algorithm in cases where the cortex is very thin or highly porous (Buie et al., 2007). In addition to limitations in measuring Ct.Th, the threshold-filter method for segmenting the cortex is severely limited for the measurement of Ct.Po as the cortical masks count large cortical pores as marrow regions (Figure 3.4). However, it should be noted that the threshold- 69

88 filter method of segmenting the cortex was not developed for the purpose of measuring cortical porosity and should therefore not be used to quantify porosity. Figure 3.4: Single slice and three-dimensional representative images of the original grayscale image (left), the segmented images from the auto-segmentation method (center) and thresholdfilter method (right). The regions designated as trabecular bone are shown in gray, cortical bone is shown in blue, and cortical pores are shown in yellow Comparison of Pre- and Postmenopausal, Normal, Osteopenic, and Osteoporotic Women To our knowledge this is the first in vivo study of Ct.Po in adult women. Results from our group comparisons are in agreement with trends observed in previous in vitro Ct.Po (Brockstedt et al., 1993) and in vivo Ct.Th (Boutroy et al., 2005) studies. To date, only one HR-pQCT study has evaluated Ct.Po in vivo (Kirmani et al., 2009). Kirmani and colleagues (2009) reported values for Ct.Po in adolescent boys and girls that were, on average, lower than those in the present study, as would be expected given the age difference of the two cohorts. However, since the threshold-filter method was used to define the cortex in the adolescent subjects, the porosity 70

89 values reported by Kirmani et al. (2009) may be underestimated due to limitations in segmenting the cortex as discussed above. Regardless of abmd category, our automated segmentation algorithm detected significant age-related differences in cortical porosity as evidenced by the higher cortical porosity at both the radius and tibia in postmenopausal women than premenopausal women. This higher porosity was due to larger pore size rather than a larger number of pores, as has also been found in histology and µct studies (Bousson et al., 2000; Cooper et al., 2007a; Stein et al., 1999b; Thomas et al., 2005). Similar differences in porosity between pre- and postmenopausal women were previously noted in vitro at the iliac crest (Brockstedt et al., 1993) and likely reflect the accelerated bone turnover that occurs with estrogen deficiency at menopause (Brockstedt et al., 1993). In addition to age-related differences in cortical porosity, our results suggest that HRpQCT is able to detect possible effects of disease on cortical porosity. Compared to women with normal abmd, postmenopausal women with osteopenia or osteoporosis had more porous cortices at both skeletal sites. In addition, HR-pQCT was able to detect differences in porosity at the distal radius and distal tibia between osteopenic and osteoporotic women. The higher porosity appeared to be due to larger pores. These findings agree with results from animal studies that reported significant increases in cortical porosity with ovariectomy (Wilson et al., 1998) and also with in vitro studies in humans that have reported disease-related increases in porosity (Christiansen et al., 1993). However, it is important to note that the present study included only a small sample of women with osteoporosis (n=9) and we were unable to make any comparisons between women with a history of fragility fracture and fracture-free women. 71

90 It is interesting to note the apparent site-specific difference in cortical porosity. On average, Ct.Po was higher at the distal tibia compared with the distal radius across all groups, and there were larger differences between the groups at the distal tibia when compared to the radius. The lower porosity the radius could be due to differences in loading environments causing higher rates of remodelling at the tibia. We also found higher variability in the porosity measures at the radius when compared to the porosity at the tibia. This is likely due to the larger average Ct.Th at the tibia, which allows for larger pores that are easier to detect. Our auto-segmentation masks were used to obtain direct measures of Ct.Th at both skeletal sites. On average, values for Ct.Th in the current study were larger in magnitude than those reported previously (Boutroy et al., 2005; Dalzell et al., 2009). For example, in the current study the average Ct.Th at the distal radius in premenopausal women was 1.06±0.18 mm whereas Boutroy and colleagues reported a lower average value for Ct.Th at the same site in premenopausal women of similar age (0.804±0.149 mm) (Boutroy et al., 2005). Similarly, Dalzell and colleagues reported a lower average Ct.Th in women of 0.78±0.23 mm at the distal radius and 0.93±0.34 mm at the distal tibia in their population based sample (Dalzell et al., 2009). The lower values for Ct.Th in these two studies are likely related to the use of the threshold-filter method to extract the cortex where undetected thin cortices contribute to a lower Ct.Th estimate. Despite differences across studies, our finding of thinner cortices at both the distal radius and tibia in the postmenopausal osteopenic and osteoporotic women compared with the premenopausal women agrees with the results of Boutroy et al. (2005) and previous reports that bone loss in postmenopausal osteoporosis occurs mainly in the cortical regions (Brown et al., 1987; Seeman & Delmas, 2006). 72

91 In contrast to our Ct.Po results, Ct.Th at both the radius and tibia did not differ between pre- and postmenopausal women with normal BMD. This finding contrasts with the HR-pQCT results of Boutroy et al. who reported significant age-related decreases in Ct.Th at both the distal radius and distal tibia in women (Boutroy et al., 2005). The differing results may again reflect differences in how the cortex was defined between the two studies. As discussed, the cortical masks generated by the filter-threshold method fail to extract the thin or highly porous cortices and exclude a significant proportion of the porosity. Measures of apparent cortical density obtained with the automatic segmentation mask showed similar trends to Ct.Po at both the distal radius and tibia. Our results agree with the findings of Boutroy et al. who reported decreased cortical density in osteoporotic and osteopenic women when compared with normal premenopausal women (Boutroy et al., 2005) and with Riggs et al. who reported increased cortical bone loss after menopause (Riggs et al., 2008). Cortical density values from the auto-segmentation were strongly correlated to cortical density values obtained with the filter-threshold method (r = 0.97); however, lower average values were obtained with the filter-threshold method. This difference is likely explained by the exclusion of cortical porosity in the filter-threshold masks as previously discussed. In summary, our fully-automated and validated segmentation method for HR-pQCT scans offers a novel approach to accurately quantify Ct.Po and Ct.Th in vivo at the distal radius and tibia. Our findings suggest that the automated segmentation method is able to detect age- and disease-related differences in cortical porosity and thickness. Further investigation is required to determine whether region-specific differences in porosity and thickness exist within the bone cross-section as has been reported in in vitro studies (Bousson et al., 2000; Thomas et al., 2005). In addition, with the application of finite element 73

92 analysis (MacNeil & Boyd, 2008a) it will be possible to examine the functional significance of Ct.Po and its contribution to bone strength, and ultimately fracture risk. Acknowledgements We would like to thank all of the subjects who participated in the study as well as Jane Allan and Bernice Love for their help with recruitment and Irene Hanley, Shannon Boucousis, and Eva Szabo for their help with scan acquisition. 74

93 CHAPTER FOUR Bone Microarchitecture and Strength in Adolescents 75

94 CHAPTER FOUR: CORTICAL POROSITY IS HIGHER IN BOYS COMPARED WITH GIRLS AT THE DISTAL RADIUS AND DISTAL TIBIA DURING PUBERTAL GROWTH: AN HR-pQCT STUDY This chapter is reprinted with permission from a manuscript published in the Journal of Bone and Mineral Research: Nishiyama KK, Macdonald HM, Moore SA, Fung T, Boyd SK, McKay HA. Cortical porosity is higher in boys compared with girls at the distal radius and distal tibia during pubertal growth: An HR-pQCT study. J Bone Miner Res 2012; 27: Author Roles: KKN contributed to the design, data analysis, interpretation, and drafting of the manuscript. SAM contributed to data acquisition and analysis. TF contributed to the statistical design and data analysis. HMM, SKB, and HAM contributed to the conception, design, and interpretation. All authors contributed to critically revising the manuscript. Abstract The aim of this study was to determine the sex- and maturity-related differences in bone microstructure and estimated bone strength at the distal radius and distal tibia in children and adolescents. We used high-resolution pqct to measure standard morphological parameters in addition to cortical porosity (Ct.Po) and estimated bone strength by finite element analysis. Participants ranged from 9-22 years (N=212 girls and N=186 boys) who were scanned annually for either one (11%) or two (89%) years at the radius and for one (15%), two (39%) or three (46%) years at the tibia. Participants were grouped by the method of Tanner into pre-pubertal, early pubertal, peri-pubertal, and post-pubertal groups. At the radius, peri- and post-pubertal girls had higher cortical density (Ct.BMD; 9.4% and 7.4%) and lower Ct.Po (-118% and -56%), 76

95 compared with peri- and post-pubertal boys (all p<0.001). Peri- and post-pubertal boys had higher trabecular bone volume ratios (p<0.001) and larger cortical cross-sectional areas (p<0.05, 0.001) when compared with girls. Based upon the load-to-strength ratio (failure load/estimated fall force), boys had lower risk of fracture than girls at every stage except during early puberty. Trends at the tibia were similar to the radius with differences between boys and girls in Ct.Po (p<0.01) and failure load (p< 0.01) at early puberty. Across pubertal groups, within sex, the most mature girls and boys had higher Ct.BMD and lower Ct.Po than their less mature peers (prepuberty) at both the radius and tibia. Girls in early, peri- and post-pubertal groups and boys in peri- and post-pubertal groups had high higher estimates of bone strength compared with their same sex pre-pubertal peers (p<0.001). These results provide insight into the sex- and maturityrelated differences in bone microstructure and estimated bone strength. 4.1 Introduction Fragility fractures of the distal radius are common injuries sustained during adolescence (Jones et al., 2002b; Landin, 1983) and are more prevalent in boys than girls (Khosla et al., 2003). The incidence of distal radius fractures peaks during the pubertal growth spurt (Bailey et al., 1989; Cooper et al., 2004b; Kramhøft & Bødtker, 1988) and epidemiological evidence suggests that the incidence of these fractures has increased by more than 30% over the past 30 years (Khosla et al., 2003). The majority of these fractures occur during play and sport; however, changes in physical activity levels and sports participation fail to explain the increasing fracture rate (Khosla et al., 2003). Alternatively, impaired bone quality during adolescence such as a deficiency in cortical bone may contribute to forearm fracture risk. As 26% of adult bone mass is 77

96 accrued during the pubertal growth spurt, impaired bone quality during growth may not only increase fracture risk during adolescence but also later in life (Bailey et al., 1999). Early dual-energy x-ray absorptiometry (DXA) studies suggest that a transient deficit in bone mass may underpin the peak incidence of forearm fractures during puberty (Bailey et al., 1999; Baxter-Jones et al., 2003; Cadogan et al., 1998). In boys and girls, rapid accrual of bone mass lags behind linear growth by approximately one year (Bailey et al., 1999). However, DXA is not capable of measuring volumetric BMD or bone microstructure elements that contribute to whole bone strength, nor can it distinguish between trabecular and cortical bone (Petit et al., 2005). Peripheral quantitative computed tomography (pqct) measures trabecular and cortical volumetric BMD, bone geometry, and provides estimates of bone strength in the growing skeleton (Kontulainen et al., 2005;, 2006; Macdonald et al., 2006;, 2005; Moyer-Mileur et al., 2001; Neu et al., 2001; Rauch et al., 2001). Previous pqct studies found that total volumetric BMD at the distal radius did not differ between girls and boys across pubertal groups (Neu et al., 2001; Rauch et al., 2001). However, there was a sex difference in cortical BMD at the distal tibia (Kontulainen et al., 2006) In addition, estimated bone strength assessed by pqct lags behind linear growth suggesting a possible transient deficit in bone strength (Rauch et al., 2001). Although pqct provides valuable information on bone structure and BMD, this technology is limited by the resolution (~400 μm in-plane), and is therefore limited in its ability to accurately measure aspects of bone microstructure such as cortical porosity (Neu et al., 2001). Most recently, high-resolution pqct (HR-pQCT) can accurately measure cortical and trabecular microstructure at the distal radius (MacNeil & Boyd, 2007a) and the distal tibia (Liu et al., 2010b). For example, at 82 μm, the resolution is sufficient to quantify cortical porosity using customized analysis tools (Buie et al., 2007; Burghardt et al., 2010b; Nishiyama et al., 2010a). 78

97 Cortical porosity may increase during puberty due to increased demands for calcium (Parfitt, 1994b). However, due to the resolution required to quantify cortical pores, cortical porosity has yet to be directly measured in vivo in children and adolescents. Porosity is negatively associated with bone strength (Wachter et al., 2001a; Yeni et al., 1997). Importantly, bone strength can now be estimated using finite element (FE) models generated from HR-pQCT images (MacNeil & Boyd, 2008a). Most HR-pQCT studies to date have focused on adult populations and provided insight into the differences in bone quality between sexes and across age groups (Dalzell et al., 2009; Khosla et al., 2006; Macdonald et al., 2011). There are very few HR-pQCT studies of adolescents and all published data are cross-sectional (Burrows et al., 2010b; Kirmani et al., 2009; Wang et al., 2010) and do not provide insight into the changes across puberty. These studies described microstructural differences during adolescence and suggested that deficiencies in cortical bone may account for the high incidence of fractures (Kirmani et al., 2009; Wang et al., 2010). HR-pQCT analyses that include standard morphological and density measurements enhanced by novel measures of cortical porosity and FE-estimated bone strength are needed to advance our understanding of bone growth and development. As the radius and the tibia serve different weight bearing and non-weight bearing roles, it is important to evaluate both to establish site specificity. These advancements may provide insight into the high incidence of distal radius fractures during the pubertal growth spurt. Therefore, the objective of this study was to use HR-pQCT to investigate differences in bone microstructure and estimated bone strength across puberty and between sexes. We also aimed to explore site-specific differences in microarchitecture and strength by evaluating both the distal radius and distal tibia. 79

98 4.2 Materials and Methods Participants Participants were healthy girls (N=212) and boys (N=186) aged 9-22 years. The older participants (14-22 yrs, N=278) were part of the Healthy Bones (HBS) III follow-up study at the University of British Columbia (Macdonald et al., 2007; MacKelvie et al., 2001; Petit et al., 2002). The younger participants (9-12 yrs, N=120) were recruited from five elementary schools in Vancouver and Richmond, British Columbia in The older participants were first scanned with HR-pQCT in 2008 at the distal tibia. In 2009 the distal radius was added to the scanning protocol for both the younger and older cohorts. Across both age cohorts at the distal tibia, there were 34 girls and 26 boys with one scan, 93 girls and 61 boys with two scans, and 85 girls and 99 boys with three scans. At the distal radius, there were 22 girls and 12 boys with one scan, and 152 girls and 131 boys with two scans (Figure 4.1). 80

99 Figure 4.1: Numbers of participants at baseline recruited from each cohort and number of followup scans analyzed for boys and girls at the distal radius and distal tibia. Our methods are described in detail elsewhere (Burrows et al., 2010b; Macdonald et al., 2007; MacKelvie et al., 2001; Petit et al., 2002). Briefly, we measured height and weight to the nearest 0.1 cm and 0.1 kg, respectively, using standard protocols. Radius length was measured as the distance from the distal medial edge of the ulna to the proximal olecranon process. Tibia length was measured as the distance from the distal edge of the medial malleolus to the tibial joint line. We assessed maturity using standard definitions based on the method of Tanner by self-reported breast stage for girls and pubic hair stage for boys (Tanner, 1978). We categorized Tanner stage 1 participants as pre-pubertal (PRE), Tanner stage 2 and 3 as early pubertal (EARLY), Tanner stage 4 as peri-pubertal (PERI), and Tanner stage 5 as post-pubertal (POST) (Macdonald et al., 2005; Wang et al., 2010). 81

100 We determined ethnicity from health history questionnaires that were previously completed by each participant s parents or guardians. As in previous studies (Burrows et al., 2010b; MacKelvie et al., 2001) participants were considered Caucasian (47%) if both parents (or all four grandparents) were born in North America or Europe, Asian (46%) if both parents (or all four grandparents) were born in China, India, Philippines, Vietnam, Korea, Taiwan or other Asian countries while 7% of the participants were of mixed or other ethnicities. All participants were healthy and none reported use of medications known to influence bone metabolism, mineralization or calcium balance. All participants and parents provided informed written consent and the University of British Columbia Clinical Research Ethics Board approved this study HR-pQCT Scan Acquisition As in previous studies (Burrows et al., 2010b), all participants were scanned using HRpQCT (XtremeCT; Scanco Medical, Brüttisellen, Switzerland) at the non-dominant radius and tibia unless there was a previous fracture at the desired site, in which case the opposite limb was scanned. A single highly trained technician positioned each participant according to the standard manufacturer protocol. For the distal radius, a two-dimensional scout view image was first acquired and the technician placed the reference line at the medial edge of the distal radius. The region of interest scanned was at a distance of 7% of the total ulnar length from the reference line (Burrows et al., 2010b) (Figure 4.2). This site ensured that we did not scan the growth plate and that we could scan the same relative region in follow-up scans. The distal tibia was scanned at the 8% site from a reference line placed at the tibial plafond (Burrows et al., 2010b) (Figure 4.2). We scanned all 82

101 participants using 60 kvp effective energy, 900 μa current, and 100 ms integration time to acquire 110 slices (approximately 9.02 mm) of the tibia and radius at an 82 μm nominal isotropic resolution. A B Figure 4.2: Scout view images of the distal radius (A) and distal tibia (B) illustrating the 7% and 8% measurement sites, respectively Image Processing and Measurements We used the standard manufacturers method to filter and binarize the HR-pQCT images (Laib et al., 1998). To segment the cortical and trabecular regions we used an automatic segmentation algorithm implemented in Image Processing Language (IPL V5.07, Scanco Medical, Brüttisellen, Switzerland) (Buie et al., 2007). Standard morphological microstructure outcomes included total bone mineral density (Tt.BMD, mg HA/cm 3 ), trabecular bone volume ratio (BV/TV), trabecular thickness (Tb.Th, mm), trabecular separation (Tb.Sp, mm), and trabecular number (Tb.N, 1/mm) (Boutroy et al., 2005; Laib et al., 1998). These measurements were validated against micro-computed tomography (MacNeil & Boyd, 2007a; Sekhon et al., 83

102 2009) and in adult populations have in vivo short-term reproducibility of <4.5% (MacNeil & Boyd, 2008b). Reproducibility in our lab is <3.8% for all parameters at the radius and tibia. In addition to standard morphological outcomes, we assessed cortical bone microstructure with our customized segmentation algorithm. Outcomes included cortical porosity (Ct.Po, %) calculated as the number of void voxels within the cortex (Buie et al., 2007; Burghardt et al., 2010b; Nishiyama et al., 2010a), cortical thickness (Ct.Th, mm) using a distance transform (Hildebrand & Rüegsigger, 1997), and cortical bone mineral density (Ct.BMD, mg HA/cm 3 ). We also calculated macrostructural parameters: cortical (Ct.Ar, mm 2 ), trabecular (Tb.Ar, mm 2 ), and total (Tt.Ar, mm 2 ) cross-sectional areas based on our customized segmentation Finite Element Analysis Using the three-dimensional HR-pQCT images, we generated finite element (FE) meshes using the voxel conversion approach (Müller & Rüegsegger, 1995; Van Rietbergen et al., 1995). We simulated uniaxial compression on each radius and tibia section up to 1% strain using a Young s modulus of 6829 MPa and Poisson s ratio of 0.3 (MacNeil & Boyd, 2008a). As previously reported, we used a custom FE solver (FAIM, Version 4.0) on a desktop workstation (Mac Pro, OSX, Version ; 2x2.8 GHz Quad-Core Intel Xenon) to solve the models (Macdonald et al., 2011). We estimated bone strength (ultimate stress, MPa) and failure load (N). In order to estimate the risk of forearm fracture, we calculated the load-to-strength ratio (Φ) (Hayes et al., 1991; Keaveny & Bouxsein, 2008) based on a subject-specific fall force that incorporates subject height (Chiu and Robinovitch, 1998). 84

103 4.2.5 Statistical Analysis We used a generalized estimating equation (GEE) model that incorporates follow-up scans as repeated measures to investigate the differences between sexes and across maturity groups. This model allows for the analysis of correlated, longitudinal data and was calculated using Equation 4.1. Equation 4.1: Generalized estimating equation to allow to analysis of correlated, longitudinal data to determine the differences across maturity groups and between sexes (Liang and Zeger, 1986). The regression parameters, β, can be obtained where Y i is the vector of measurements for the i th subject, u i is the vector of means, and V is the covariance matrix. The dependent variables were bone microstructure measurements and FE estimates of bone strength while the fixed factors were maturity and sex. We used limb length and height as covariates in the model. If there was a significant interaction effect detected between maturity sex, we also evaluated the simple effects of maturity and sex. We performed pairwise comparisons between girls and boys within each maturity group. In order to determine changes across puberty, we performed comparisons within each sex between the least mature group (PRE) and the other maturity groups. We used SPSS Statistics (IBM, Version 19.0; Somers, NY) for all analyses and a Bonferroni correction to account for multiple comparisons. 85

104 4.3 Results Participant Descriptives Descriptive characteristics of the participants by maturity groups are provided in Table 4.1. Within each sex, as expected, more mature boys and girls were significantly taller, heavier, and had longer radii and tibiae compared with their less mature peers. Between sexes, there were significant differences in height, weight, and limb length in the PERI and POST groups Comparison Between Sexes Within Maturity Categories For all parameters with the exception of Tb.Sp and Tb.N at both sites, and Tb.Th at the distal tibia, there was a statistically significant sex maturity interaction (Table 4.2 and Table 4.3). At the distal radius, the sex difference in bone density and microstructure was most evident in children in the PERI and POST groups (Table 4.2 and Table 4.3). In these groups, girls had higher Ct.BMD (p<0.001) and lower Ct.Po (p<0.001; Figure 4.3) than boys. In contrast, boys in the PERI and POST groups had higher BV/TV (both p<0.001) and Tb.Th (PERI p<0.05, POST p<0.001) compared with girls. Boys in the PERI and POST groups had a larger Ct.Ar compared with girls in the same groups (p<0.05 and 0.001, respectively). Estimated distal radius bone strength (ultimate stress), which accounts for cross-sectional area, was not significantly different between girls and boys at any stage. In contrast, estimated failure load was higher in boys than girls in the PERI and POST groups (both p<0.001). Based on the calculated load-tostrength ratio, boys had a significantly lower risk of radial fracture at every stage, with the exception the EARLY group (Table 4.4). 86

105 Figure 4.3: Plots of cortical density (Ct.BMD), cortical porosity (Ct.Po), cortical area (Ct.Ar), and failure load at the distal radius for girls and boys by puberty group. Error bars represent SE. a. p<0.001, b. p<0.01, c. p<0.05: significant difference between girls and boys within the same puberty group. d. p<0.001 e. p<0.01: significant difference between puberty group and the PRE group within sex. All p-values are after Bonferroni correction. We observed similar sex differences in bone outcomes at the distal tibia in children in the PERI and POST groups; however, sex differences in some variables were also present in the EARLY group (Table 4.3). With the exception of the PRE group, girls had lower Ct.Po compared with boys (p<0.01, 0.001, 0.001, EARLY to POST respectively, Figure 4.4). Boys had a larger Tt.Ar in the EARLY, PERI and POST groups and Ct.Ar in all groups compared with 87

106 girls. While ultimate stress at the distal tibia was not significantly different between boys and girls, failure load estimates at the distal tibia were higher in boys compared with girls in each pubertal group (p<0.01, 0.01, 0.001, 0.001, PRE to POST respectively). Figure 4.4: Plots of cortical density (Ct.BMD) and cortical porosity (Ct.Po) at the distal tibia for girls and boys by puberty group. Error bars represent SE. a. p<0.001, b. p<0.01: significant difference between girls and boys within the same puberty group. d. p<0.001 e. p<0.01: significant difference between puberty group and the PRE group within sex. All p-values are after Bonferroni correction Comparison Across Maturity Categories Within Sex At the distal radius, both girls and boys in all groups had higher Tb.Ar, Tt.Ar, failure load, and load-to-strength ratios than their same sex peers in the PRE group (Table 4.2 and Table 4.4). For both sexes, BV/TV was not different in any of the groups compared with the PRE group. For boys and girls, Ct.BMD and Tt.BMD were higher in the PERI and POST groups compared with the same sex PRE group (all p<0.001). Girls had lower Ct.Po in the PERI and 88

107 POST groups compared with the PRE girls and POST boys had lower Ct.Po compared with PRE boys (all p<0.001). Ultimate stress estimates from FE analysis were higher in girls in all groups and in boys in the PERI and POST groups compared with the same sex PRE group (EARLY girls: p<0.05, remainder: p<0.001). For both girls and boys at the distal tibia, Tt.Ar, Ct.Ar, and failure load were higher in all groups compared with the PRE group (Table 4.3 and Table 4.4). Similar to the distal radius, BV/TV was not significantly different across pubertal groups in boys or girls. For girls, Ct.Th was higher in all groups compared with the PRE group (all p<0.001), while boys had higher Ct.Th only in the PERI and POST groups compared with boys in the PRE group (both p<0.001). In boys, Ct.Po was not significantly different across pubertal groups. In girls, Ct.Po was lower in PERI (p<0.01) and POST (p<0.001) girls compared with PRE girls. Finite element estimates of ultimate stress were higher in girls in all groups compared with PRE girls (all p<0.001), and higher in PERI and POST boys compared with PRE boys (both p<0.001). 89

108 Table 4.1: Participant descriptives by puberty group for girls and boys [Mean (SD)]. Note: All p values after Bonferroni correction. a. p<0.001, b. p<0.01, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex. 90

109 Table 4.2: Bone microstructure at the distal radius across puberty groups for girls and boys [Mean (SD)]. All p-values after Bonferroni correction. a. p<0.001, b. p<0.01, c. p<0.05, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex. 91

110 Table 4.3: Bone microstructure at the distal tibia across puberty groups for girls and boys [Mean (SD)]. All p-values after Bonferroni correction. a. p<0.001, b. p<0.01, c. p<0.05, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex. 92

111 Table 4.4: Finite element estimated bone strength parameters at the distal radius and distal tibia across puberty groups for girls and boys [Mean (SD)]. All p-values after Bonferroni correction. a. p<0.001, b. p<0.01, c. p<0.05, significant difference between girls and boys within the same puberty group. d. p<0.001, e. p<0.01, f. p<0.05, significant difference between puberty group and the PRE group within sex. 93

112 4.4 Discussion Our findings characterize the many sex- and maturity-related differences in growing bone microstructure and strength at the weight bearing tibia and the non-weight bearing radius. We extend the existing paediatric bone quality literature by supplementing standard HR-pQCT morphological analysis of the growing radius and tibia with novel measures of cortical porosity, estimated bone strength and failure load. At both skeletal sites we report significant differences in cortical and trabecular bone macro- and microstructure between girls and boys across pubertal stages (Figure 4.5). In particular, as others (Wang et al., 2010) and we (Burrows et al., 2010a) previously reported, bone size was consistently larger in boys compared with girls and girls cortices were more dense compared with boys. However, we also report girls less porous cortices compared with boys in the more mature groups (EARLY, PERI, POST). Thicker trabeculae underpin the significant differences in trabecular outcomes in PERI and POST boys compared with PERI and POST girls. The relevance of our findings relates to the transient deficit in cortical bone that may explain the high incidence of distal radius fractures in boys and girls during the pubertal growth spurt (Cooper et al., 2004b; Jones et al., 2002a). 94

113 Figure 4.5: A schematic representation of differences in total bone size and cortical bone density for girls (G) and boys (B) across puberty (assessed using the method of Tanner (T)). For our purposes we defined Tanner stage 1 as pre-puberty (PRE), Tanner stages 2 and 3 as early puberty (EARLY), Tanner stage 4 as peri-puberty (PERI) and Tanner stage 5 as post-puberty (POST). Significant differences between girls and boys are shown for finite element estimated failure load (where boys values exceed girls after early puberty) and cortical porosity (Ct.Po) where boys values exceed girls after pre puberty. (Diagram not exact scale) To date, one HR-pQCT study examined the distal radius and distal tibia during growth (Wang et al., 2010) while another only examined the distal radius (Kirmani et al., 2009). While the small sample size in the study by Wang and colleagues (Wang et al., 2010) limited their 95

114 ability to detect significant maturity- and sex-related differences, the trends in radius trabecular microstructure, Ct.BMD, and Tt.Ar and all tibia measurements were similar to our findings. However, unlike Wang et al., we did not observe lower Ct.Th in EARLY boys compared with PRE boys. This discrepancy may be due to differences in the methods used to segment the cortical and trabecular regions. In the present study we used a validated automatic segmentation method (Buie et al., 2007; Nishiyama et al., 2010a) whereas Wang and colleagues defined cortical bone as five consecutive bone voxels. However, this approach may be undermined by high Ct.Po along the endocortical surface in boys during early puberty. This would result in lower values for Ct.Th and Ct.Ar. Our findings align with the only other study to use HR-pQCT to assess the distal radius of adolescents (Kirmani et al., 2009). Kirmani et al. found that Ct.Th did not differ between Tanner stage 1 and Tanner stage 3 boys (Kirmani et al., 2009). As reported in a previous pqct study (Rauch et al., 2001), we found that Ct.Th was significantly higher at the distal radius in boys than girls in the POST group, only. Across maturity groups, we found a difference in cortical bone structure and strength between girls and boys. Although Ct.BMD was similar between PRE and EARLY girls and boys, PERI and POST girls had less porous and more dense cortices compared with boys at both the radius and tibia. Parfitt thought this sex difference in Ct.BMD may be due to increased demand for calcium during the pubertal growth spurt, a demand met by increased intracortical bone turnover (Parfitt, 1994b). During this transient process, increased bone turnover would, in turn, lead to increased cortical porosity. Boys experience a greater height velocity (4.9 vs. 2.9 cm/year) on average, and at peak (Bailey et al., 1989) and a longer growth spurt (Tanner et al., 1976), compared with girls. Thus, the higher values for Ct.Po and lower Ct.BMD we observed in boys may well be explained by boys greater height velocity. 96

115 In their recent cross-sectional study Kirmani and colleagues calculated a cortical porosity index using the volume of cortical pores divided by the total cortical volume (Kirmani et al., 2009) using the Gaussian filter-threshold method to segment the cortical and trabecular regions, which was different from our approach. They observed significantly greater values for this index in boys at bone-age group 4 and girls at bone-age group 3 (which approximates our PERI and EARLY groups, respectively). Despite girls advantage at the cortex, boys are conferred a substantial structural advantage by their having larger bones (represented by total cross-sectional area). Specifically, small additions of bone to the periosteal surface confer increased resistance to compressive forces and exponentially increased resistance to bending forces (Ruff and Hayes, 1982). In order to provide better insight into the peak incidence of forearm fractures during puberty, we used finite element analysis to estimate bone strength. Since ultimate stress is adjusted by bone cross-sectional area it was not significantly different between boys and girls at any maturity time point. However, when we predicted failure load and the load-to-strength ratio (Keaveny & Bouxsein, 2008) (which incorporates estimated fall force), boys were significantly higher than girls in their same maturity category except for the EARLY group. Despite adjusting for limb length and height, the apparent paradox of no difference in estimated failure load between early pubertal girls and boys, rather than higher failure loads in girls, may speak to the inability of FE analysis to adequately capture the high porosity that leads to a transient weakness in growing bone. One solution may be to, in future, impose other loading conditions in addition to compressive loading to assess bone strength. The apparent site-specificity for sex differences in bone microstructure and FE estimates of bone strength (Pritchett, 1991;, 1992) at this site is indicated by higher failure loads compared 97

116 with girls. The tibia had higher failure loads than the radius as well as higher estimates of sizeadjusted ultimate stress likely due to the weight-bearing nature of the site. We note that our study has limitations. Due to the northern location of the participants (latitude: 49 degrees), it is possible that some participants may suffer from a degree of hypovitaminosis D. Due to similar exposure across participants we do not expect this to influence our results. However, the absolute values may differ from a population in a more southern geographical location. Also, we did not account for the ethnic diversity of our cohort in our analysis. There are documented differences in bone mass (Burrows et al., 2009) between Asians and Caucasians; however, as our cohort also included children of mixed ethnicity we chose to include participants of all ethnicities in our analysis. In addition, we accept that based on the maturational indices used we were unable to align boys and girls at exactly the same maturational time point. Longer term prospective studies that span the pubertal growth spurt and that allow more direct and comparable measures of maturity (i.e., age at peak height velocity) are needed to more specifically ascertain the differences in bone quality between boys and girls. It is also important to note that our customized method used to quantify cortical porosity (Nishiyama et al., 2010a) cannot detect pores smaller than 82 μm. These smaller pores may be more prevalent in children and adolescents. In addition, our estimates of bone strength and failure load were based on uniaxial compression, which may not be an accurate representation of the forces applied in an actual fall on an outstretched hand. Thus, future research should incorporate a greater variety of loading conditions. Finally, load-to-strength ratios based on fall force calculations have not been validated in adolescent populations; however, this estimate of fracture risk was used in recent studies of young adults (Dalzell et al., 2009; Macdonald et al., 2011) and adolescents (Kirmani et al., 2009). 98

117 Despite these limitations, our findings provide a comprehensive description of maturityrelated differences in bone quality and strength in girls and boys. It also encourages research to delve further into the high incidence of distal radius fractures during the pubertal growth spurt. Prospective studies of bone microstructure comparing children who have sustained a fracture with their non-fractured age or maturity-matched counterparts would serve to further elucidate the concept of transient bone fragility. Acknowledgements We wish to thank to the students, staff and parents in the Vancouver and Richmond School Districts for their continued support and participation in the Healthy Bones III Study. We also thank the staff at the Centre for Hip Health and Mobility, in particular Danmei Liu, for their assistance with data collection and ongoing support of this research. 99

118 CHAPTER FIVE Distinguishing Fragility Fractures With HR-pQCT 100

119 CHAPTER FIVE: WOMEN WITH PREVIOUS FRAGILITY FRACTURES CAN BE CLASSIFIED BASED ON BONE MICROARCHITECTURE AND FINITE ELEMENT ANALYSIS MEASURED WITH HR-pQCT This chapter is based on a manuscript that has been accepted for publication in Osteoporosis International: Nishiyama KK, Macdonald HM, Hanley DA, Boyd SK. Women with previous fragility fractures can be classified based on bone microarchitecture and finite element analysis measured with HR-pQCT. Osteoporosis International. Author Roles: KKN contributed to the design, data analysis, interpretation, and drafting of the manuscript. HMM, DAH, SKB contributed to the conception, design, and interpretation. All authors critically revised the manuscript. Abstract Areal bone mineral density (abmd) is the primary measurement used to assess osteoporosis and fracture risk; however, does not take into account bone microarchitecture, which also contributes to bone strength. Thus, our objective was to determine if bone microarchitecture measured with HR-pQCT and FE estimates of bone strength could classify women with and without low-trauma fractures. We used HR-pQCT to assess bone microarchitecture at the distal radius and tibia in 44 postmenopausal women with a history of low-trauma fracture and 88 age-matched controls from the Calgary cohort of the CaMos study. We estimated bone strength using FE analysis and simulated distal radius abmd from the HRpQCT scans. Femoral neck and lumbar spine abmd were measured with DXA. We used support vector machines and a 10-fold cross-validation to classify the fracture cases and controls, and 101

120 determined accuracy. The combination of HR-pQCT measures of microarchitecture and FE estimates of bone strength had the highest area under the receiver operating characteristic curve (AUC) of 0.82 when classifying forearm fractures compared to an AUC of 0.71 from DXAderived abmd of the forearm and 0.63 from femoral neck and spine DXA. For all fracture types, FE estimates of bone strength at the forearm alone resulted in an AUC of Models based on HR-pQCT measurements of bone microarchitecture and estimates of bone strength performed better than DXA-derived abmd at classifying women with and without prior fracture. In future, these models may improve the prediction of individuals at risk of low-trauma fracture. 5.1 Introduction Areal bone mineral density (abmd) measured by dual energy x-ray absorptiometry (DXA) is a significant predictor of fracture risk (Black et al., 1992); however, half of fractures occur in women who would not be classified as osteoporotic by abmd (Schuit et al., 2004). The mechanical strength of bone is dependent on both density and microarchitecture (van der Linden and Weinans, 2007) and measurements of cortical and trabecular microarchitecture are independent of abmd (Sornay-Rendu et al., 2007). Thus, incorporating bone structural information with density measurements could potentially improve assessment of fracture risk. High-resolution peripheral quantitative computed tomography (HR-pQCT) is an emerging technology capable of measuring cortical and trabecular microarchitecture at the distal radius and distal tibia (Boutroy et al., 2005; MacNeil & Boyd, 2007a). The resolution of HRpQCT also permits application of finite element (FE) analysis to the 3D scans to estimate bone strength. These FE models are highly correlated with bone strength measured directly with mechanical testing (MacNeil & Boyd, 2008a) as they incorporate both the material and 102

121 architectural components of bone strength. Previous HR-pQCT studies examined the ability of individual bone microarchitectural (Boutroy et al., 2008; Melton et al., 2007; Vico et al., 2008; Vilayphiou et al., 2010) and FE analysis parameters (Boutroy et al., 2008; Melton et al., 2007; Vilayphiou et al., 2010) or a combination of parameters (Boutroy et al., 2008; Vilayphiou et al., 2010) to discriminate between postmenopausal women with and without a history of low-trauma fracture. These studies showed that bone structure and strength contribute to forearm fracture risk independently of abmd (Melton et al., 2007). It has been an important first step to establish the association of microarchitectural parameters and FE outcomes with low-trauma fractures (Vilayphiou et al., 2010). This naturally leads us to determine whether it is possible to classify people with low-trauma fracture. Machine learning methods are statistical tools used to recognize patterns in data sets. These methods are often used for the purpose of classification since they can be trained based on known cases, and then tested on new cases. Machine learning methods can be used to incorporate all HR-pQCT and FE parameters in order to classify those with and without fracture. This is in contrast to statistical approaches such as principal component analysis, which reduce the data into combinations of parameters that account for the majority of the variance. Atkinson and colleagues recently used a machine learning method, gradient boosting machines (GBM), and found that fracture prediction improved when all possible bone density, geometry and microstructural parameters obtained with central QCT and HR-pQCT were included in the GBM model when compared with DXA parameters alone (Atkinson et al., 2012). Support vector machines (SVM) are an alternative machine learning method that can be used to classify individuals with and without previous fracture by maximizing the separation between groups. 103

122 SVM models offer several advantages over other machine learning methods, including greater stability and minimal requirements for parameter tuning. Thus, our objective was to use SVM models to incorporate bone microarchitectural parameters and FE estimates of bone strength to determine if we can classify women with and without previous low trauma fractures. 5.2 Materials and Methods Participants Participants in this study were postmenopausal women who were members of the Calgary, AB cohort of the Canadian Multicentre Osteoporosis Study (CaMos) (Macdonald et al., 2011; Nishiyama et al., 2010b). Briefly, CaMos is a 10-year prospective population-based study in which participants were recruited using a stratified random sampling technique from nine centres across Canada (Kreiger et al., 1999). At the 10-year follow-up, we invited individuals from the Calgary CaMos cohort to participate in an HR-pQCT sub-study. From this cohort (N=442; ages years), we identified postmenopausal women who sustained a low-trauma fracture (excluding finger, face, and toes) during the 10 years of study follow-up (n=44). Fractures were identified by self-report during scheduled interviews at years 3, 5 and 10 and by yearly postal questionnaires through year 9. Additional information regarding circumstances surrounding the fracture and fracture site were gathered via a structured telephone interview (Berger et al., 2009). Low-trauma fractures were those that occurred without trauma or as a result of a fall from standing height or less. We randomly matched two participants without fracture by age (±1.0 year) to each fracture case (n=88). We determined height, weight, bisphosphonate and corticosteroid use (> 3 months), and menopause status from the interviewer-administered CaMos questionnaire. In addition, as part of the standard CaMos assessment, trained technicians 104

123 measured femoral neck (FN) and lumbar spine (LS) areal bone mineral density (abmd, g/cm 2 ) using dual energy x-ray absorptiometry (DXA, Hologic QDR4500, Bedford, Massachusetts). The Conjoint Health Research Ethics Board at the University of Calgary approved all procedures HR-pQCT Scan Acquisition and Measurements As we describe in detail elsewhere (Macdonald et al., 2011; Nishiyama et al., 2010b), we scanned all participants using HR-pQCT (XtremeCT; Scanco Medical, Brüttisellen, Switzerland) at the non-dominant radius and left tibia unless there was a previous fracture at the desired site, in which case we scanned the opposite limb. One of two highly trained operators acquired and analyzed all scans according to the manufacturer s standard in vivo protocol. We scanned all participants using 60 kvp effective energy, 1000 μa current, and 100 ms integration time to acquire 110 slices (approximately 9.02 mm) of the radius and tibia at an 82 μm nominal isotropic resolution. Scans were manually scored for motion on a scale of 0 (no motion) to 4 (significant blurring of the periosteal surface, discontinuities in the cortical shell, or streaking in the soft tissue) and scans scored as a 4 were excluded from this analysis (n=2). We used the manufacturer s standard method to filter and binarize the HR-pQCT images (Laib et al., 1998) and assessed all standard HR-pQCT morphological microstructure outcomes (Boutroy et al., 2005). These measurements were previously validated against micro-computed tomography (MacNeil & Boyd, 2007a; Sekhon et al., 2009) and have in vivo short-term reproducibility of <4.5% in adult populations (MacNeil & Boyd, 2008b). Reproducibility in our lab is <3.8% for all parameters at the radius and tibia. 105

124 To segment the cortical and trabecular regions, we used an automatic segmentation algorithm implemented in Image Processing Language (IPL V5.07, Scanco Medical, Brüttisellen, Switzerland) (Buie et al., 2007). Based on this segmentation, we calculated macrostructural parameters: cortical (Ct.Ar, mm 2 ), trabecular (Tb.Ar, mm 2 ), and total (Tt.Ar, mm 2 ) cross-sectional areas. In addition, we calculated cortical porosity (Ct.Po, %) as the number of void voxels within the cortex (Buie et al., 2007; Nishiyama et al., 2010b) and measured direct cortical thickness (Ct.Th, mm), and cortical bone mineral density (Ct.BMD, mg HA/cm 3 ). We acquired FN and LS abmd by DXA, but did not acquire ultra distal radius (UDR) DXA scans for the participants. Therefore, we implemented the method developed by Burghardt and colleagues (Burghardt et al., 2009) to simulate DXA-derived abmd: we aligned the HRpQCT scans in the same configuration as a DXA scan and obtained a calibrated projection image from which we determined abmd. This method is strongly correlated with DXA abmd (R 2 = 0.82) and is highly reproducible with a root mean-squared coefficient of variation of 1.1% Finite Element Analysis From the HR-pQCT images we generated finite element (FE) meshes using the voxel conversion approach (Müller & Rüegsegger, 1995; Van Rietbergen et al., 1995). Each voxel was converted to an 8-node, 82 μm 3, hexahedral element. The meshes generated resulted in approximately 1 million elements for the radius scans and 2.5 million elements for the tibia scans. We simulated uniaxial compression on each radius and tibia model up to 1% strain using a Young s modulus of 6829 MPa and Poisson s ratio of 0.3 (MacNeil & Boyd, 2008a) as previously reported (Macdonald et al., 2011; Nishiyama et al., 2010b). We used a custom FE solver (FAIM, Version 4.0; Numerics88 Solutions, Calgary, Canada) on a desktop workstation 106

125 (Mac Pro, OSX, Version ; 2x2.8 GHz Quad-Core Intel Xenon) to estimate bone strength (ultimate stress, MPa) based on the relationship determined by MacNeil and Boyd (2008). We also determined the percentage of load carried by the cortex based on the elements labeled as cortical bone from the automatic segmentation at the most distal and most proximal slices. In order to estimate the risk of forearm fracture, we calculated the load-to-strength ratio (Φ) (Chiu & Robinovitch, 1998; Hayes et al., 1991; Keaveny & Bouxsein, 2008). This ratio represents the estimated fall force on the outstretched arm divided by the estimated failure load from the FE analysis. The theoretical fracture threshold occurs when Φ >1.0 (Chiu & Robinovitch, 1998; Hayes et al., 1991; Keaveny & Bouxsein, 2008) Statistical Analysis We used support vector machines (SVM) (Vapnik, 1982) with a weighted radial basis kernel function to classify women with and without low trauma fracture. In the first analysis we examined only low trauma forearm fractures and the second analysis we included all low trauma fractures. We used Student s t-test to compare outcome variables between the fracture and control groups. For each analysis we generated SVM models based on standardized measurements of 1) all HR-pQCT, FE, and DXA parameters, 2) all HR-pQCT and FE parameters, 3) FE parameters alone, 4) UDR, LS, and FN abmd and 5) LS and FN abmd. For the forearm fractures we used only the distal radius measurements and for all types of fracture we used the distal radius and distal tibia measurements. To gauge the accuracy of the SVM results and to avoid over-fitting to our specific dataset we used a 10-fold cross validation scheme. This method randomly divides the participants into 10 groups, trains the model on nine of these groups, and tests the model on the remaining group. This is repeated 10 times so that 107

126 each group is used as the test group and the accuracy reported is the average of all iterations. The results of this validation are presented as the areas under the curve (AUC) for receiver operating characteristic (ROC) curves. WEKA (Version 3.7; University of Waikato; Hamilton, NZ) was used for the SVM analyses (Hall et al., 2009), and SPSS Statistics (Version 19.0; IBM; Somers, NY) for all other analyses. 5.3 Results Participant Characteristics and Bone Outcome Variables Descriptive characteristics of the participants are provided in Table 5.1. The fracture and non-fracture groups were similar with respect to age, height, weight, BMI, and use of corticosteroids. Use of bisphosphonates was higher among women with a history of forearm fracture compared with non-fracture controls (p=0.028). The median time between fracture and scan date was 4.5 years (Interquartile Range: years). The majority of fractures were forearm fractures (n=14) and lower leg (n=14), followed by upper arm (n=9), spine (n=5), hip (n=4), upper leg (n=4), and ribs (n=2). Eight women reported two low-trauma fractures; we classified these women based on their first fracture. We present the HR-pQCT and FE variables for the radius and tibia in Table 5.2 and Table 5.3, respectively, and DXA parameters in Table 5.4. Significant differences (p<0.05) between the forearm fracture group and non-fracture controls were found at the radius for all parameters with the exception of Tb.Th, Ct.Po, and Ct.Ar. When all types of fractures were pooled all parameters were significantly different with the exception of Tb.Th, Ct.Po, Ct.BMD, Ct.Th, and area measurements. At the distal tibia, Tb.N, Tb.Sp, Ct.BMD, Ct.Po, Tt.Ar, Tb.Ar, and the percentage of the load carried by the cortical region at the distal and proximal slices were 108

127 not significantly different between forearm fracture cases and controls. For all fracture types pooled at the distal tibia, Ct.BMD, Ct.Po, Tt.Ar, Tb.Ar, and the percentage of the load carried by the cortical region were not significantly different between fracture cases and controls. The percentage of women who had load-to-strength ratios greater than 1.0 and therefore theoretically at greater risk of fracture was 28.6% for those with forearm fractures and 10.7% for their corresponding non-fracture controls; however, the mean of the entire groups were still both below 1.0. For women with all fracture types 25.0% exceeded a ratio of 1.0 while 13.6% of their corresponding controls exceeded the threshold. The forearm fracture women had significantly lower LS abmd and UDR abmd compared with the controls. The women with fractures at all sites had significantly lower FN and LS abmd from their corresponding controls SVM Classification The accuracy, sensitivity, specificity, and ROC-AUC s from the cross validation of the SVM models are summarized in Table 5.5. For the forearm fractures and controls, the highest accuracy (83.3%) and AUC (0.82) was achieved using the HR-pQCT measurements and FE estimates of bone strength together. This was followed by the combination of HR-pQCT, FE, and DXA parameters (Accuracy: 81.0%, AUC: 0.80) and by FE parameters alone (Accuracy: 78.6%, AUC: 0.80). Using LS, FN, and UDR abmd produced an accuracy of 71.4% and an AUC of 0.71, while using only LS and FN abmd had the worst performance (Accuracy: 59.5%, AUC: 0.63) classifying the women. When considering low trauma fractures at all sites, the SVM models were not as effective at classifying the fracture and control participants compared with the focused forearm fracture analysis. The FE parameters alone had the highest accuracy (69.7%) and had the highest AUC (0.69). The combination of HR-pQCT and FE measurements had an 109

128 accuracy of 68.9% and an AUC of 0.67 while the HR-pQCT, FEA and DXA abmd also had an accuracy of 68.9% and an AUC of

129 Table 5.1: Descriptive characteristics of the postmenopausal women with forearm fractures and fractures at all sites, and their corresponding age-matched controls. 111

130 Table 5.2: Bone microarchitecture and finite element parameters [mean (SD)] for the fracture and control groups at the distal radius. 112

131 Table 5.3: Bone microarchitecture and finite element parameters [mean (SD)] for the fracture and control groups at the distal tibia. 113

132 Table 5.4: DXA parameters [mean (SD)] for the fracture and control groups. Table 5.5: SVM classification results for the forearm fractures and for fractures at all sites. Five different models were used for each fracture group. The first was HR-pQCT, FE, and all DXA parameters, the second was HR-pQCT and FE parameters, the third was FE parameters alone, the fourth was all DXA parameters including simulated ultra-distal radius (UDR), and the last was lumbar spine (LS) and femoral neck (FN) DXA only. Accuracy, sensitivity, specificity, and the receiver operating characteristic area under the curve (ROC-AUC) are shown for each model. 114

133 5.4 Discussion In this study we successfully demonstrated that a machine-learning technique based on support vector machine (SVM) model could classify postmenopausal women with and without a history of low trauma fracture based on HR-pQCT input. The SVM model incorporating HRpQCT bone microarchitectural and density, as well as FE estimates of bone strength at the distal radius, outperformed DXA-derived abmd which is the clinical gold standard for osteoporosis screening and classification of forearm fracture. The SVM method, incorporating all macro- and microstructural outcomes as well as estimates of bone strength, is a promising tool for classification of women with and without previous low-trauma fracture and could potentially be an important clinical basis for fracture risk assessment. Figure 5.1: Representative distal radius scans of a participant with a low trauma forearm fracture (left) and an age-matched non-fracture control (right). Support vector machines are a type of machine learning tool that recognize patterns in data and map data to a higher dimension to create nonlinear boundaries between the cases (Cortes & Vapnik, 1995), and notably they are stable and require minimal tuning of parameters 115

134 (Valyon and Horváth, 2003). This allows SVMs to work well in many applications and reduces over-fitting to the specific dataset. Other machine learning methods such as gradient boosting machines (GBM) may provide higher accuracy of classification, but are computationally expensive and, most importantly, require extensive tuning of parameters. SVMs have an important advantage in that they find a global, unique solution unlike other methods such as artificial neural networks that may find local minima (Suykens et al., 2002). This advantage of SVMs potentially makes this approach more robust for application to diverse clinical datasets. To date, only one other study has used machine-learning techniques to classify postmenopausal women with and without a history of low trauma (Atkinson et al., 2012). Similar to our findings, Atkinson et al. reported more accurate classification when bone density, geometry and microstructural outcomes obtained with additional imaging modalities (QCT and HR-pQCT) were included in their GBM model in addition to standard DXA-derived abmd indicating the importance of microarchitectural parameters. In addition, whereas not all bone outcomes were significantly different between groups when compared individually, small contributions of each variable combined into a single model allowed for better fracture discrimination. Atkinson et al. reported higher AUC s for predicting forearm fracture cohorts compared with our current study, which was possibly due to their larger sample size (n= 99 forearm fractures) (Atkinson et al., 2012). While our study is smaller, we used a 10-fold cross validation scheme, which gives an excellent indication of how our model would perform in practice on a new dataset. If our model was trained and subsequently tested on the same current dataset, it would likely show higher AUC s; however, we would be at higher risk of being overfit to our specific dataset. 116

135 Previous studies used principal component analysis (PCA) and logistic regression models to determine if HR-pQCT measures of bone microarchitecture and FE-estimates of bone strength could discriminate between postmenopausal women with and without previous fracture (Boutroy et al., 2008; Vico et al., 2008; Vilayphiou et al., 2010). These analyses differ from our approach in that they attempt to determine the individual parameters (or principal components of parameters) that best discriminate fracture cases from controls, whereas the goal of the SVM is to perform classification. Despite the different analytic approaches and goals, results of these studies consistently indicate that HR-pQCT measurements of bone microarchitecture and FE analysis estimates of bone strength outperform DXA measured abmd when discriminating between postmenopausal women with and without a history of low trauma fracture. Consistent with our findings, Vilayphiou et al. showed that FE estimated bone stiffness had high contributions to the principal components (Vilayphiou et al., 2010). The relatively similar performance between our model using HR-pQCT, DXA, and FE parameters and the FE parameters alone indicate that the FE measurements are providing an excellent summary measurement of bone quality. Clinical DXA scans of the forearm are rare; therefore, we included models based on DXA measurements with and without UDR abmd. The model with only LS and FN abmd represents a typical clinical scenario and, as expected, resulted in the weakest classification as shown in Table 5.5. Based on the model that included UDR, FN, and LS abmd, it is important to include a scan of the distal radius to classify forearm fractures. Our data may actually overestimate the importance of a UDR scan because our simulated UDR abmd measurements were focused on precisely the area that is most likely to fracture, whereas a real UDR DXA would cover a broader region, and may be less sensitive to classifying forearm fracture. It is not 117

136 surprising that models including UDR abmd more accurately classify forearm fractures. However, HR-pQCT measurements still outperformed all DXA measurements when classifying forearm fractures. We found that the SVMs were better at discriminating low trauma forearm fractures compared with low trauma fractures at other sites. This is likely attributable to site-specificity since the HR-pQCT scans obtained at the distal radius were used to classify the forearm fractures. The difference in discriminative ability may also be due to the large variation in fracture types in our study. We had few participants (n=6) with fractures at major osteoporotic sites (i.e. hip and spine). It is possible these fractures would be better predicted by the SVM s because women who suffer these types of fractures tend to be at more advanced disease stages compared with women who sustain forearm fractures (Harvey et al., 2009). A larger cohort of fracture cases would strengthen our understanding of the potential for SVMs to classify fractures. However, our current study indicates that there is excellent potential for this method in classification studies of all low trauma fractures. A limitation of our study is that it is a cross-sectional design and a relatively small sample size, and the acquisition of HR-pQCT scans were done retrospectively after the fractures were sustained. Despite these weaknesses, our study provides strong support for SVM combined with HR-pQCT to classify fractures better than by the clinical gold standard of DXA abmd, even at the UDR. It would be ideal to expand this research with a larger fracture cohort, and to perform the study prospectively, and work in this area is currently underway at our laboratory and at other sites worldwide. Another limitation of our study is that the FE estimates are based on a single uniaxial compression test. While that is currently the gold standard for FE applied to HRpQCT, it may be advantageous to expand the number of simulated FE tests for a more 118

137 comprehensive assessment of bone strength, and hence better predict the strength of the bone during a fall on an outstretched arm. In addition, FE models were computed using a homogenous Young s modulus for each element. Strength estimates using this method are highly correlated with experimental estimates of bone strength; however, FE measurements may be affected by mineralization differences. In future, it may be important to account for such differences using scaled modulus values based on density measurements. The application of SVMs to HR-pQCT data shows great promise to classify postmenopausal women with and without a history of low trauma fracture. Our results suggest that this is an accurate classification method for forearm fractures, and has potential to be applicable for classification of fractures at other sites. This is the first study to apply SVM s to HR-pQCT data and is strengthened by incorporating cross validation for testing, and by including FE estimates of bone strength. Acknowledgements We wish to acknowledge the ongoing efforts of the national Canadian Multicentre Osteoporosis Study (CaMos). We would also like to thank Ms. Irene Hanley and Ms. Shannon Boucousis for their assistance with scan acquisition, and we are grateful to all of the participants who volunteered for the study. 119

138 CHAPTER SIX Proximal Femur Strength Estimated by Finite Element Analysis 120

139 CHAPTER SIX: PROXIMAL FEMUR BONE STRENGTH ESTIMATED BY A COMPUTATIONALLY FAST FINITE ELEMENT ANALYSIS IN A SIDEWAYS FALL CONFIGURATION This chapter is based on a manuscript that has been submitted for publication: Nishiyama KK, Gilchrist S, Guy P, Cripton P, Boyd SK. Proximal femur bone strength estimated by a computationally fast finite element analysis in a sideways fall configuration. Author Roles: KKN contributed to the design, experimental data acquisition, data analysis, interpretation, and drafting of the manuscript. SG contributed to experimental data acquisition and interpretation. PG and PC contributed to the interpretation. SKB contributed to the conception, design, and interpretation. All authors contributed to critically revising the manuscript. Abstract Finite element (FE) analysis based on quantitative computed tomography (QCT) images is an emerging tool to estimate bone strength in a specific patient or specimen; however, it is limited by the computational power required and the associated time required to generate and solve the models. Thus, our objective was to develop a fast, validated method to estimate whole bone structural stiffness and failure load in addition to a sensitivity analysis of varying boundary conditions. We performed QCT scans on twenty fresh-frozen proximal femurs (Age: 77 ± 13 years) and mechanically tested the femurs in a configuration that simulated a sideways fall on the hip. We used custom software to generate the FE models with boundary conditions corresponding to the mechanical tests and solved the linear models to estimate bone 121

140 structural stiffness and estimated failure load. For the sensitivity analysis, we varied the internal rotation angle of the femoral neck from -30 to 45 at 15 intervals and estimated structural stiffness at each angle. We found both the FE estimates of structural stiffness (R 2 = 0.89, p<0.01) and failure load (R 2 = 0.81, p<0.01) to be in high agreement with the values found by mechanical testing. An important advantage of these methods was that the models of approximately 500,000 elements took less than 11 minutes to solve. In this study we developed and validated a method to quickly and accurately estimate proximal femur structural stiffness and failure load using QCT-driven FE methods. 6.1 Introduction While some hip fractures occur during normal activity (Parker and Twemlow, 1997), the impact forces encountered during a sideways fall are believed to be the main factor in over 90% of hip fractures (Grisso et al., 1991; Nevitt et al., 1989). Areal bone mineral density (abmd) measured by dual energy x-ray absorptiometry (DXA) has been shown to be a significant predictor of fracture risk (Kanis, 2002) on a population level; however, 50% of fractures occur in people who, according to the World Health Organization abmd criteria, would not be classified as osteopenic let alone as osteoporotic (Stone et al., 2003). Incorporating quantitative computed tomography (QCT) images and finite element (FE) analysis to estimate bone strength on a persubject basis is a promising technique to improve the individual assessment of fracture risk. This non-invasive, subject-specific technique overcomes some of the inherent limitations of DXA (Bolotin, 2007) by incorporating volumetric measurements of density with structural information. 122

141 Finite element analysis and QCT images have been used to estimate whole bone stiffness and failure load in both standing and falling load configurations (Bessho et al., 2009;, 2007; Cody et al., 1999; Dragomir-Daescu et al., 2011; Duchemin et al., 2008; Keyak et al., 1998;, 2001; Koivumäki et al., 2012; Lochmüller et al., 2002; Lotz et al., 1991a;, 1991b). These models have been in very good agreement with experimental mechanical testing of cadaver specimens (Bessho et al., 2007; Cody et al., 1999; Dragomir-Daescu et al., 2011; Duchemin et al., 2008; Keyak et al., 1998; Koivumäki et al., 2012) and have better agreement with femoral strength compared with abmd by DXA or volumetric BMD by QCT (Cody et al., 1999). One of the main limitations of these models is their time-consuming and computationally expensive nature making them difficult to apply to clinical studies. Linear FE models have taken up to 6-8 hours (Cody et al., 1999) while non-linear models can take approximately 10 hours each (Bessho et al., 2009; Keyak, 2001). The loading configuration of the proximal femur has a large influence on the experimentally measured (Pinilla et al., 1996) and estimated whole bone strength (Bessho et al., 2009; Keyak et al., 2001; Wakao et al., 2009). Pinilla and colleagues found that a 30 change in loading angle decreased the failure load by 24% (Pinilla et al., 1996). Since it is unknown how a specific person will fall (i.e. directly sideways, slightly turned forward or backward), and the large effect of loading direction on structural stiffness is well documented, it is important to determine the sensitivity of estimates of bone strength to various loading configurations. Again, applying this to a clinical population is limited by the computational time and power required to solve the models in various configurations. Another challenge with developing proximal femur FE models is the segmentation of the cortical and trabecular regions. In some regions, the thickness of the cortical shell can approach the resolution of the QCT system, leading to partial 123

142 volume effects that can cause errors in the segmentation. An accurate segmentation is required to assign appropriate material properties to the cortical and trabecular regions (Hangartner, 2007) and for accurate post-processing to determine the contributions of the cortical and trabecular regions to estimated whole bone strength. Therefore, our primary objective was to develop a fast, validated method to estimate bone stiffness and failure load. Our secondary objective was to determine the sensitivity of boundary conditions and the contributions of the cortical and trabecular regions to the overall bone stiffness. 6.2 Materials and Methods Specimens We collected twenty fresh-frozen proximal femurs (5 male, 15 female; average age: 77 ± 13 years) from the Gross Anatomy Lab at the University of Calgary. No specific requirements were used for the selection and no medical history was available so it was not known if previous bone diseases were present. Before scanning and mechanical testing, we cleaned the specimens of soft tissue, kept them frozen in saline soaked gauze, and thawed them overnight. The Conjoint Health Research Ethics Board at the University of Calgary approved all procedures CT Scan Acquisition Prior to scanning, all specimens were secured in saline filled tube in order to simulate the attenuation of soft tissue. The tubes were placed on top of a hydroxyapatite calibration phantom (B-MAS200, Kyoto Kagaku, Japan) that contained five hydroxyapatite rods (0, 50, 100, 150, 200 mg/cm 3 ) and extended the length of the scan region. We scanned all specimens with CT (GE 124

143 Discovery CT750HD, GE Healthcare; 120 kvp, 60 mas, 512x512 matrix size) and reconstructed the images with a slice thickness of mm and an in-plane resolution of mm x mm. Total femoral neck (FN) abmd by DXA (Hologic QDR4500, Hologic; Bedford, MA) was also measured for all specimens Mechanical Testing We performed the mechanical testing in a configuration designed to simulate a sideways fall on the hip (de Bakker et al., 2009; Courtney et al., 1994; Eckstein et al., 2004; Manske et al., 2006; Pulkkinen et al., 2006). The testing apparatus positioned the femoral shaft at an angle of 10 from horizontal and the femoral neck internally rotated to 15. The femoral shaft was free to translate in the sagittal plane and free to rotate around the anterior-posterior axis at the distal end. The femoral head was also free to translate in the sagittal plane but constrained in the direction of loading. The femoral head and the greater trochanter were both embedded in polymethylmethacrylate (PMMA; Fastray, Bosworth Company, Skokie, Il.) caps to prevent local crushing and improve load distribution. The mechanical testing apparatus and specimen was placed within a materials testing system (Instron 8874; Instron Corp; Canton, MA) with a 25 kn rated load cell (Sensor Data M211-11; Sterling Heights, MI) (Figure 6.1). A data acquisition card (National Instruments PCI- 6040E; Austin TX) signal conditioned with a 10 khz hardware cutoff filter and sampled at 20 khz recorded the load and displacement. We used a 2000 N/V output resulting in a bit level resolution of 4.9 N/bit for the load and 5 mm/v resulting in a mm/bit resolution for the displacement. First, we applied a preload of 100 N followed by a hold of 2 seconds. Next we applied a constant displacement at a rate of 2 mm/second. From the force-displacement curve, 125

144 we determined the stiffness based on the slope of the linear region of the curve and failure load as the maximal load encountered during the test (Pistoia et al., 2002). Figure 6.1: Photograph of the testing apparatus and specimen positioned in the mechanical testing device (top) and corresponding schematic of the finite element model boundary conditions where a 1 mm displacement is applied to the PMMA cap (represented in white) on the greater trochanter (bottom). 126

145 6.2.4 Image Processing To find the periosteal and endosteal surfaces we used a semi-automatic contouring method (Stradwin 4.3; Cambridge, UK) (Treece et al., 2010). This software produces an unbiased cortical thickness measurement to 0.3 mm (Treece et al., 2010) and also outputs the cortical and trabecular surfaces. The surfaces were used to classify the cortical and trabecular regions to the nearest voxel. In-house software developed using the Visualization Toolkit (VTK 5.6; Kitware Inc.; Clifton Park, NY) calibrated the QCT measured density values based on the calibration phantom and rescaled the images using cubic interpolation, as previously suggested (McErlain et al., 2012), to 1.0 mm isotropic voxels Finite Element Analysis To create the FE model, we converted each of the voxels in the image to an 8-node hexahedral element and assigned a Poisson s ratio of 0.3 (Varghese et al., 2011). The calibrated density values (ρ ash, mg/cm 3 ) were converted to Young s modulus values (E, GPa) using the 2.29 relationship found by Keller (E = ρ ash Equation 6.1) (Keller, 1994). E = ρ ash 2.29 Equation 6.1: Density (ρ ash, mg/cm 3 ) to modulus (E, MPa) relationship determined by Keller (1994) for cortical and trabecular bone. 127

146 The boundary conditions in the FE model mimicked the mechanical testing conditions that represent a sideways fall on the hip (Figure 6.1). This included the embedded shaft and PMMA caps on both the femoral head and the greater trochanter. We assigned elements representing the PMMA caps a Young s modulus of 2500 MPa and a Poisson s ratio of 0.3 (Lewis, 1997) and fixed the surface nodes on the femoral head PMMA cap in the loading direction, but they were free to move in the transverse directions. A single line of nodes on the end of the femoral shaft holder was fixed to simulate the pivot point in the mechanical test. Finally, a 1 mm displacement was applied to the surface nodes on the trochanter PMMA cap. We solved the linear FE models using FAIM (v5.4, Numerics88 Solutions; Calgary, Canada) installed on a desktop workstation (Mac Pro, OSX, Version Apple Inc., Cupertino, CA, USA; 2 x 2.4 GHz Quad-Core Intel Xenon, 8GB RAM). We estimated stiffness based on the reaction force and displacement data and the failure load based on the failure criteria developed by Pistoia and colleagues (Pistoia et al., 2002). This method uses the effective strain distribution in the model to estimate when failure occurs. Based on a linear scaling of the tissue strains, we estimated failure load when 7% of the voxels exceeded 0.9% strain, which gave the most consistent bias in the estimates. Based on the segmentation of the cortical and trabecular regions, we determined the percentage of the load carried by the cortical and trabecular regions across the femoral neck. In the majority of previous studies an internal rotation angle of 15 was used. In order to determine the sensitivity of the loading direction to bone stiffness and estimated strength, we adjusted the internal rotation angle of the femoral neck to -30, -15, 0, 15, 30, and 45. The resultant stiffness of each specimen was compared at the various simulated loading directions. 128

147 6.2.6 Statistical Analysis To validate the FE outcomes, we used linear regression to assess the relationship between the experimental and FE estimated stiffness and failure load. In addition, we used Bland-Altman plots to compare the bias of the FE estimated outcomes to the experimental values (Bland & Altman, 1986). The FE estimated stiffness and failure load were compared across the various loading configurations using a repeated measures analysis of variance and, after a Bonferoni adjustment, considered a p-value of <0.05 to be significant. All analysis used SPSS Statistics (IBM, Version 19.0; Somers, NY). 6.3 Results Descriptive characteristics and femoral neck abmd measured by DXA for the femoral specimens are provided in Table 6.1. The FE estimates of bone stiffness were strongly predictive of the experimental mechanical testing values (R 2 = 0.89, p<0.01) (Figure 6.2). Failure load predicted by the FE models was also highly predictive of the experimental failure loads determined from the mechanical testing (R 2 = 0.81, p<0.01) (Figure 6.2). The FE estimates of stiffness and failure load both tended to underestimate the experimental values as depicted in the Bland-Altman plots in Figure 6.2; however, there were no significant correlations in the Bland- Altman plots indicating there was a consistent bias. 129

148 Table 6.1: Age, sex, weight, height and total femoral neck (FN) abmd measured by DXA for the cadaver proximal femur specimens. Specimen # Age (years) Side Sex Weight (kg) Height (cm) FN abmd (g/cm 2 ) 1 55 L F R F L M R M L F R F L M R M L F R F L F N/A N/A L F R F L F R F L M L F N/A N/A R F N/A N/A L F N/A R F

149 Figure 6.2: Comparison of experimental versus the finite element estimates of stiffness (A) and failure load (B). Bland-Altman plots depict the mean of the experimental and FE estimated stiffness (C) versus the difference between the two values and the corresponding plot for failure load (D). The horizontal lines indicate the mean values and 95% confidence intervals. For the sensitivity analysis, the significant differences (p<0.05) in FE estimated stiffness after Bonferroni adjustment are shown in Figure 6.3. Compared with estimated stiffness at 45, all other estimates were significantly higher. Estimates at -15, 0, and 15 were higher than the stiffness estimated at 30 and estimates at -15 and 0 were higher than at 15. Finally, the 131

150 stiffness estimate at -15 was significantly higher than at -30 (Figure 6.3). Through the femoral neck region 68.4% ± 4.9% of the load was carried by the cortical region in the 15 neck rotation position. The mean time to solve the FE models was 10.9 ± 2.0 minutes based on an average of 543,134 ± 41,379 elements per model. From the 20 models we generated, smallest model was 483,826 elements, which took 8.1 minutes to solve. The largest model we had was 624,711 elements and took 14.8 minutes to solve. Figure 6.3: The overall FE estimated structural stiffness in the loading direction with the femoral neck internally rotated -30, -15, 0, 15, 30, and 45. After a Bonferroni adjustment, significantly (p<0.05) higher stiffness values are indicated: a) compared with 45, b) compared with 30, c) compared with 15, and d) compared with -30. Error bars represent standard error. 132

151 6.4 Discussion In this study we developed a method to quickly generate and solve large FE models of the proximal femur simulating a sideways fall. We validated the models using mechanical testing and found good agreement between the estimated stiffness (R 2 = 0.89) and failure load (R 2 = 0.81) and the values measured by mechanical testing. The minimal user intervention generating the models and computational speed of the model solver is promising to become a clinically applicable tool for the estimation of bone strength and fracture risk. In order to make the method faster and more clinically relevant we used two approaches. First, we used a voxel conversion approach to generate the FE models (Müller & Rüegsegger, 1995; Van Rietbergen et al., 1995), which compared to traditional meshing approaches, saves time and eliminates user bias. In addition, this method directly utilizes the QCT scan data and does not set any arbitrary values such as a minimum cortical thickness as done in previous studies (Bessho et al., 2007; Koivumäki et al., 2012). Second, we used an element-by-element, preconditioned conjugant gradient (PCG) solver (Hughes et al., 1987; Van Rietbergen et al., 1995). While hardware advances play an important role in decreasing computation time, the speed of this FE solver can also be attributed to careful implementation of the PCG solver. This takes advantage of the fact that the global stiffness matrix is never computed and the computations can be distributed to multiple cores in a desktop computer. We chose to test the proximal femurs in a falling condition to mimic the forces that occur in over 90% of hip fractures (Grisso et al., 1991; Nevitt et al., 1989). While many studies have used a loading configuration to simulate standing (Bessho et al., 2007; Cody et al., 1999; Duchemin et al., 2008; Keyak, 2001), recent studies used this more clinically relevant loading configuration (Dragomir-Daescu et al., 2011; Keyak et al., 1998; Koivumäki et al., 2012). 133

152 Dragomir-Daescu et al. used step-wise linear models and Koivumäki et al. used bilinear elastoplastic models in a falling configuration to estimate bone stiffness and strength (Dragomir- Daescu et al., 2011; Koivumäki et al., 2012). Dragomir-Daescu and colleagues reported a R 2 = 0.87 between experimental stiffness and FE estimated stiffness and an R 2 = 0.85 for ultimate load; however, even with a novel mesh implementation, their models took approximately 10 hours to solve (Dragomir-Daescu et al., 2011). Koivumäki et al. reported an R 2 = 0.87 for fracture load and while they did not specify the average time to solve the models, they indicated that that it was a time consuming process and not sufficient for clinical use (Koivumäki et al., 2012). Keyak and colleagues also implemented a linear FE solver and achieved very high correlations between experimentally measured fracture load and estimated fracture load (R 2 = 0.90) (Keyak et al., 1998). Like our current study, the results from Keyak et al. implies that that accuracy of non-linear analysis may not outweigh the extensive computation time using current software and hardware. The relationships used for the conversion of density values measured by QCT to elastic moduli have been variable between studies (Helgason et al., 2008), but are an essential step in the modeling process. For our study we chose to use the relationship determined by Keller (Keller, 1994) since it was most relevant to our study. First, the relationship determined by Keller was based on human femoral bone, and studies have shown that density-to-modulus relationships are site-specific (Goulet et al., 1994; Morgan et al., 2003). While it would be possible to have a different density-to-modulus relationship for cortical and trabecular bone, even with advanced segmentation techniques, the resolution of QCT and the consequent partial volume effects makes this distinction difficult. In addition, Keller s relationship spans a wider 134

153 density range (0.028 to g/cm 3 ) than most other studies thus providing data for both cortical and trabecular bone (Keller, 1994). We also performed a sensitivity analysis and found that the boundary conditions have a significant effect on the estimated bone stiffness. The highest stiffness values were found at 0 and -15 of internal rotation to the femoral neck representing a fall on the hip slightly rolled forward and the lowest found at 45 of internal rotation representing a fall on the hip where the impact is posterolateral. Our results are consistent with trends found by previous FE studies (Bessho et al., 2009; Ford et al., 1996; Keyak et al., 2001; Wakao et al., 2009) and mechanical testing studies (Pinilla et al., 1996) that examined the effects of boundary conditions. Since there is a large effect of boundary conditions and the impact direction of a specific fall is unknown, it may be important to consider various real-world boundary condition configurations when attempting to estimate femoral bone strength in an in vivo clinical population. The percentage of the load carried by the cortical region in the femoral neck was on average 68.4%. This measure may be important when comparing osteoporotic and healthy patients as was suggested by Verhulp et al. where an osteoporotic proximal femur model carried a larger proportion of the load on the cortex compared to a healthy proximal femur (Verhulp et al., 2008). There are limitations to the experimental testing and FE modeling in our study that should be noted. For our mechanical testing, we used a relatively low loading rate of 2 mm/second that does not reflect the dynamic loading rates that may be encountered during a sideways fall (Robinovitch et al., 1995a;, 1995b); however, our loading rates reflect those that were used to determine the density-to-modulus relationships. For the FE models, we used the method described by Pistoia and colleagues based on the effective strain distribution (Pistoia et al., 2002) to estimate failure load. While this method was optimized for FE models of the distal 135

154 radius and does not distinguish between compressive and tensile strains, it resulted in a very high agreement with the actual failure loads determined from the mechanical tests (R 2 = 0.81). Further studies may be needed to choose the optimal effective strain and percentage of failed voxels for the failure load estimation. Finally, the resolution of the CT scans is limited and partial volume effects may affect the assignment of material properties in the model. In order to minimize these effects, we used an accurate segmentation of the cortical and trabecular regions (Treece et al., 2010) and in the FE models used a relatively small voxel size (1 mm 3 ). In this study we developed a fast method that requires minimal user intervention to accurately estimate subject-specific proximal femur stiffness and failure load. The accuracy of the current methods is comparable to other studies that have used more complex, non-linear FE solvers, but has the major advantage of, on average, taking less than 12 minutes per femur model to solve. Due to the nature of the solver, there is even more potential for faster computation times by using multicore computers or graphics processor units. Future work will focus on incorporation non-linear analysis as well as anisotropic material properties. With increased automation and continuous improvements in computational modeling, QCT-based FE analysis for non-invasive estimates of bone strengths is a promising tool on the clinical horizon. Acknowledgements We wish to thank the Gross Anatomy Lab at the University of Calgary for collection and preparation of the specimens. In addition we would like to thank Dr. Masako Ito for use of the calibration phantom and Drs. David McErlain and Clara Sandino for assistance with specimen preparation and scanning. Finally, we would like to acknowledge the technical programming assistance from Mr. Eric Nodwell. 136

155 CHAPTER SEVEN Fracture Classification with QCT and Finite Element Analysis 137

156 CHAPTER SEVEN: CLASSIFICATION OF WOMEN WITH AND WITHOUT HIP FRACTURE BASED ON QUANTITATIVE COMPUTED TOMOGRAPHY AND FINITE ELEMENT ANALYSIS This chapter is based on a manuscript that is in preparation for publication: Nishiyama KK, Ito M, Harada A, Boyd SK. Classification of women with and without hip fracture based on quantitative computed tomography and finite element analysis. Author Roles: KKN contributed to the conception, design, data analysis, interpretation, and drafting of the manuscript. MI contributed to conception, data acquisition, and interpretation. Ah contributed to design and data acquisition. SKB contributed to the conception, design, and interpretation. All authors contributed to critically revising the manuscript. Abstract Areal bone mineral density (abmd) is the current standard for assessing fracture risk; however, many fractures occur in people not identified as osteoporotic by abmd. Finite element (FE) analysis based on quantitative computed tomography (QCT) images takes into account both bone material and structural properties to provide subject-specific estimates of bone strength. Thus our objective was to determine if FE estimates of bone strength could classify age-matched women with and without hip fracture. Twenty women with femoral neck fracture and 15 women with trochanteric fractures along with 35 corresponding age-matched controls were scanned with QCT at the hip. Since it is unknown how a specific subject will fall, FE analysis was used to estimate bone stiffness and bone failure load under loading configurations with femoral neck internal rotation angles ranging from -30 to 45 with 15 intervals. Support vector machine (SVM) models and a 10-fold cross-validation scheme were used to classify the subjects with and 138

157 without fracture. High accuracy was achieved when classifying the women with and without fracture both when the fracture types were pooled (84.3%) and when they were analyzed separately by neck fracture (85.0%) and trochanteric fracture (80.0%). Areas under the receiver operating characteristic curve were also high for the pooled fractures (0.84), neck fractures (0.85), and trochanteric fractures (0.80). While larger, prospective studies are needed, these results demonstrate the ability of FE analysis using multiple loading configurations, combined with SVMs, to provide subject-specific classification for hip fracture. 7.1 Introduction Finite element (FE) analysis based on images from quantitative computed tomography (QCT) has tremendous potential to estimate bone strength and ultimately, improve fracture prediction. Unlike areal bone mineral density (abmd) measured by dual x-ray absorptiometry (DXA), this method has the advantage of taking into account both material and structural properties of bone. Much of the research combining FE analysis with QCT images has focused on the proximal femur (Amin et al., 2011; Bessho et al., 2009;, 2007; Cody et al., 1999; Dragomir-Daescu et al., 2011; Duchemin et al., 2008; Grassi et al., 2012; Keaveny et al., 2010; Keyak, 2001; Keyak et al., 1998;, 2011; Koivumäki et al., 2012; Orwoll et al., 2009; Viceconti et al., 2004; Wakao et al., 2009) since fractures at this site result in high mortality, morbidity, and health care costs (Cummings and Melton, 2002; Melton III, 2003). To date, these methods have been validated in cadaver studies using mechanical testing (Bessho et al., 2007; Cody et al., 1999; Dragomir-Daescu et al., 2011; Duchemin et al., 2008; Keyak et al., 1998; Koivumäki et al., 2012). 139

158 Recent studies have combined FE analysis and QCT images to determine age-related differences in bone strength (Keaveny et al., 2010; Lang et al., 2012) as well as estimate bone strength in women treated with teriparatide or alendronate (Keaveny et al., 2012). In addition, investigators have examined associations between FE estimates of bone strength and hip fractures (Amin et al., 2011; Keyak et al., 2011; Orwoll et al., 2009) and have found that FE estimates of bone strength are strongly associated with fracture in both men (Amin et al., 2011; Keyak et al., 2011; Orwoll et al., 2009) and women (Amin et al., 2011; Keyak et al., 2011). These analyses have simulated the forces encountered in a standing loading configuration (Keyak et al., 2011) or a configuration simulating a sideways fall (Amin et al., 2011; Keyak et al., 2011; Orwoll et al., 2009). Over 90% of proximal femur fractures occur during a fall (Grisso et al., 1991; Nevitt et al., 1989); however, the exact configuration in which an individual will fall (i.e. rolled forward or backward) is not possible to predict. Therefore, estimating bone strength at more than one loading configuration may improve fracture prediction. Until recently, performing several FE analyses on the same bone model has been restricted by the computational resources required, but advances in computational strategies and hardware are now making it possible to pursue this promising approach. Classification of fracture subjects on the basis of QCT and FE analysis is a challenge. One option is support vector machine (SVM) models, which is a machine learning technique that can be used to recognize patterns in datasets (Vapnik, 1982). Rather than simply finding an association between measured parameters and fracture outcomes, SVMs can be used to classify new cases on an individual basis. They are becoming more popular due to their stability and minimal requirements for parameter tuning (Valyon & Horváth, 2003). These SVMs have successfully been used to classify subjects as osteoporotic based on dental radiographs (Kavitha 140

159 et al., 2012) and to discriminate subjects with and without fracture based on radiographs of the wrist (Lee et al., 2008). To our knowledge, SVMs have not been used to classify subjects with and without fracture based on FE estimates of bone strength. Thus, the objective of this study was to estimate bone strength in a set range of loading configurations, and to determine if we could use SVMs to accurately classify women with and without hip fractures. 7.2 Materials and Methods Participants Participants who suffered hip fractures were recruited from the Aichi and Nagasaki prefectures in Japan, in addition to fracture-free, age-matched controls as previously described in detail elsewhere (Ito et al., 2010). There were a total of 35 women with fractures, of which 20 women (average age: 80.1 ± 4.5 years) had a femoral neck fracture and 15 women (average age: 82.6 ± 5.0 years) had trochanteric fractures, as classified by a radiologist. Women without fracture (N=35, average age: 79.9 ± 3.1 years) were randomly matched by age (± 5.0 years) to the women with femoral neck and trochanteric fractures. Both the Nagasaki University Hospital (Japan) and the National Center for Geriatric and Gerontology (Japan) Internal Review Boards approved all procedures, and all participants provided written informed consent CT Scan Acquisition The women were either scanned with QCT at Nagasaki University Hospital (Aquilion 16; Toshiba Medical Systems; Tokyo, Japan) or the National Center for Geriatrics and Gerontology Hospital (SOMATOM Cardiac 64; Siemens Medical Solutions; Forchheim, Germany). Women 141

160 who suffered a fracture were scanned between the time of fracture and the time of surgery. All of the women were scanned while lying on a hydroxyapatite calibration phantom (B-MAS200; Kyoto Kagaku; Japan) that contained five hydroxyapatite rods (0, 50, 100, 150, 200 mg/cm 3 ) and extended the length of the scan region. The scans at Nagasaki University Hospital were reconstructed with an in-plane resolution of mm 2 and the scans at the National Center for Geriatrics and Gerontology Hospital were reconstructed at an in-plane resolution of mm 2. Scans from both sites had a slice thickness of 0.5 mm and were scanned using 120 kvp and 250 mas. There was a high linear correlation (r = , p<0.001, y = 1.03x mg/cm 3 ) between the density values measured by the scanners at the two sites based on a quality assurance phantom (QCT Pro QA Phantom; Mindways; San Francisco, USA) Image Processing In the women without a fracture, the left femur was analyzed. In the case of the women with fracture, the fractured-free femur was analyzed. Volumetric bone mineral density (vbmd, g/cm 3 ) was measured using commercial software (QCT Pro; Mindways; San Francisco, USA) at the femoral neck. A semi-automatic contouring method (Stradwin 4.3; Cambridge, UK) (Treece et al., 2010) was used to identify the periosteal and endosteal surfaces and classify the cortical and trabecular regions to the nearest voxel in the QCT images. We used in-house software developed using the Visualization Toolkit (VTK 5.6; Kitware Inc.; Clifton Park, USA) to calibrate the CT measured density values based on the calibration phantom and rescaled the images using cubic interpolation to 1.0 mm isotropic voxels (McErlain et al., 2012). 142

161 7.2.4 Finite Element Analysis To create the FE model, we used a validated method developed in-house and described in detail in Chapter 6. Briefly, the calibrated density values (mg HA/cm 3 ) were converted to Young s modulus values (E, GPa) using the Keller s relationship (Keller, 1994) and all elements were assigned a Poisson s ratio of 0.3 (Varghese et al., 2011). The loading conditions were designed to mimic those that would be expected in a typical sideways fall on the hip. This involved applying the load to the greater trochanter, fixing the femoral head in the loading direction, and positioning the femoral shaft 10 from horizontal. We performed the FE analysis with a set range of incremented angles of internal rotation to the femoral neck (-30, -15, 0, 15, 30, and 45 ) to simulate multiple fall configurations (Figure 7.1). We solved the linear FE models using FAIM (v5.4, Numerics88 Solutions; Calgary, Canada) installed on a desktop workstation (Mac Pro, OSX, Version Apple Inc., Cupertino, CA, USA; 2 x 2.4 GHz Quad-Core Intel Xenon, 8GB RAM). Whole bone stiffness was estimated based on the reaction force and displacement data and the failure load was estimated based on the failure criteria developed by Pistoia and colleagues (Pistoia et al., 2002). While it is to be acknowledged that these failure criterion are limited due to their basis on a linear FE model, the study in the previous chapter has revealed that this simple approach shows high agreement (R 2 = 0.81) with experimental fracture results. 143

162 Figure 7.1: Schematic of the angles of rotation for the femoral neck used for estimating the bone strength. A left femur is shown with the femoral neck internal rotation angle ranging from -30 to 45 at 15 intervals Statistical Analysis To determine differences in the age, weight, height, vbmd, estimated stiffness and failure load between the women with fracture and their corresponding controls, we used an analysis of variance and adjusted for multiple comparisons using a Bonferroni correction. We used support vector machines (SVM) with a weighted radial basis kernel function to classify women with and without fracture. Models were generated for (1) the femoral neck fracture group, (2) the trochanteric fracture group, and (3) both groups pooled. To determine which parameters were the best predictors of fracture, we compared three models, each based on different parameters. The first model for each group included only vbmd to classify the women. The second model included estimated bone stiffness and failure load from the FE analysis performed at all internal rotation angles. Finally, the last model for each group included vbmd in addition to FE estimates of bone strength. To gauge the accuracy of the SVM results, we used a 10-fold cross-validation scheme, as is commonly used (McLachlan et al., 2004). The results of this validation are presented as the areas under the curve (AUC) from the receiver operating 144

163 characteristic (ROC) curves. WEKA (Version 3.7; University of Waikato; Hamilton, NZ) was used for the SVM analyses (Hall et al., 2009), and SPSS Statistics (Version 19.0; IBM; Somers, USA) for all other analyses. 7.3 Results Descriptive characteristics for the women are provided in Table 7.1. The fracture and fracture-free groups were not different with respect to age, height, and weight. The mean time to solve the FE models was 47.4 ± 9.8 seconds, and models had an average of 85,984 ± 18,029 elements. There were significant differences in vbmd, estimated bone stiffness, and estimated failure load between the women who suffered either type of fracture compared to the agematched, fracture-free women (Table 7.2). There were no significant differences between the women with trochanteric fracture compared to those with a femoral neck fracture in vbmd, stiffness, or failure load (all p-values >0.18). The accuracy, sensitivity, specificity, and ROC-AUCs from the cross-validation of the SVM models are summarized in Table 7.3. When fractures were pooled together, the highest accuracy (88.6%) and AUC (0.89) was achieved using the model containing both vbmd and FE measurements. The sensitivity of this classification was 90.0% and the specificity was 80.0%. The next best SVM model used FE measurements only (Accuracy: 84.3%, AUC: 0.84), and finally the model with vbmd only (Accuracy: 72.9%, AUC: 0.73). When only considering women with a femoral neck fracture and their corresponding controls, an accuracy of 85.0%, AUC of 0.85, and sensitivity and specificity of 88.9% and 81.8%, respectively was achieved for both the model with FE measurements and the model with 145

164 FE measurements and vbmd. In comparison, using vbmd only produced an accuracy of 82.5% and AUC of When classifying the women with trochanteric fractures from their corresponding fracture-free controls the best classification was using only the FE measurements resulting in an accuracy of 80.0% and AUC of 0.80 (Sensitivity: 90.9%, Specificity: 73.7%). 146

165 Table 7.1: Descriptive characteristics as Mean ± SD for all fractures versus controls then broken down into trochanteric fractures versus controls, and neck fractures versus controls. P-values represent differences between the fracture groups and their corresponding controls 147

166 Table 7.2: FE estimates of stiffness, failure load and QCT-vBMD. Results are grouped by all fractures versus controls then broken down by trochanteric fractures versus controls, and neck fractures versus controls. a. p<0.001, b. p<0.01, c. p<0.05 between the fracture groups and their corresponding controls after Bonferroni adjustment. 148

167 Table 7.3: Accuracy, sensitivity, specificity, and receiver operating characteristic areas under the curve (ROC-AUC) for the SVM models. The first are based on both types of fractures pooled together, the second on trochanteric fractures, and the third only on femoral neck fractures. 149