3D Reconstruction of Long Bones Utilising Magnetic Resonance Imaging (MRI)

Thesis submitted by Kanchana Rathnayaka Rathnayaka Mudiyanselage MBBS

This thesis is submitted in fulfilment of the requirements for the degree of Doctor of Philosophy

Institute of Health and Biomedical Innovation School of Engineering Systems Faculty of Built Environment and Engineering

Queensland University of Technology Brisbane, Australia 2011

Abstract

Abstract The design of pre-contoured fracture fixation implants (plates and nails) that correctly fit the anatomy of a patient utilises 3D models of long bones with accurate geometric representation. 3D data is usually available from computed tomography (CT) scans of human cadavers that generally represent the above 60 year old age group. Thus, despite the fact that half of the seriously injured population comes from the 30 year age group and below, virtually no data exists from these younger age groups to inform the design of implants that optimally fit patients from these groups. Hence, relevant bone data from these age groups is required. The current gold standard for acquiring such data–CT–involves ionising radiation and cannot be used to scan healthy human volunteers. Magnetic resonance imaging (MRI) has been shown to be a potential alternative in the previous studies conducted using small bones (tarsal bones) and parts of the long bones. However, in order to use MRI effectively for 3D reconstruction of human long bones, further validations using long bones and appropriate reference standards are required. Accurate reconstruction of 3D models from CT or MRI data sets requires an accurate image segmentation method. Currently available sophisticated segmentation methods involve complex programming and mathematics that researchers are not trained to perform. Therefore, an accurate but relatively simple segmentation method is required for segmentation of CT and MRI data. Furthermore, some of the limitations of 1.5T MRI such as very long scanning times and poor contrast in articular regions can potentially be reduced by using higher field 3T MRI imaging. However, a quantification of the signal to noise ratio (SNR) gain at the bone - soft tissue interface should be performed; this is not reported in the literature. As MRI scanning of long bones has very long scanning times, the acquired images are more prone to motion artefacts due to random movements of the subject‟s limbs. One of the artefacts observed is the step artefact that is believed to occur from the random movements of the volunteer during a scan. This needs to be corrected before the models can be used for implant design. As the first aim, this study investigated two segmentation methods: intensity thresholding and Canny edge detection as accurate but simple segmentation methods for segmentation of MRI and CT data. The second aim was to investigate the III

Abstract

usability of MRI as a radiation free imaging alternative to CT for reconstruction of 3D models of long bones. The third aim was to use 3T MRI to improve the poor contrast in articular regions and long scanning times of current MRI. The fourth and final aim was to minimise the step artefact using 3D modelling techniques. The segmentation methods were investigated using CT scans of five ovine femora. The single level thresholding was performed using a visually selected threshold level to segment the complete femur. For multilevel thresholding, multiple threshold levels calculated from the threshold selection method were used for the proximal, diaphyseal and distal regions of the femur. Canny edge detection was used by delineating the outer and inner contour of 2D images and then combining them to generate the 3D model. Models generated from these methods were compared to the reference standard generated using the mechanical contact scans of the denuded bone. The second aim was achieved using CT and MRI scans of five ovine femora and segmenting them using the multilevel threshold method. A surface geometric comparison was conducted between CT based, MRI based and reference models. To quantitatively compare the 1.5T images to the 3T MRI images, the right lower limbs of five healthy volunteers were scanned using scanners from the same manufacturer. The images obtained using the identical protocols were compared by means of SNR and contrast to noise ratio (CNR) of muscle, bone marrow and bone. In order to correct the step artefact in the final 3D models, the step was simulated in five ovine femora scanned with a 3T MRI scanner. The step was corrected using the iterative closest point (ICP) algorithm based aligning method. The present study demonstrated that the multi-threshold approach in combination with the threshold selection method can generate 3D models from long bones with an average deviation of 0.18 mm. The same was 0.24 mm of the single threshold method. There was a significant statistical difference between the accuracy of models generated by the two methods. In comparison, the Canny edge detection method generated average deviation of 0.20 mm. MRI based models exhibited 0.23 mm average deviation in comparison to the 0.18 mm average deviation of CT based models. The differences were not statistically significant. 3T MRI improved the contrast in the bone–muscle interfaces of most anatomical regions of femora and tibiae, potentially improving the inaccuracies conferred by poor contrast of the articular regions. Using the robust ICP algorithm to align the 3D surfaces, the step IV

Abstract

artefact that occurred by the volunteer moving the leg was corrected, generating errors of 0.32 ± 0.02 mm when compared with the reference standard. The study concludes that magnetic resonance imaging, together with simple multilevel thresholding segmentation, is able to produce 3D models of long bones with accurate geometric representations. The method is, therefore, a potential alternative to the current gold standard CT imaging.

V

Keywords

Keywords Magnetic resonance imaging Computed tomography Image segmentation 3D models Long bones Thresholding Edge detection Multi thresholding Higher field MRI Musculoskeletal MRI Motion artefacts Validation

VI

Contents

Contents Abstract

................................................................................................................. III

Keywords

................................................................................................................. VI

List of figures ............................................................................................................. XIII List of tables ................................................................................................................ XV Publications, presentations and awards .................................................................... XVI Authorship

.............................................................................................................. XIX

Acknowledgement ...................................................................................................... XXI Abbreviations ............................................................................................................ XXII Chapter 1.

Introduction .............................................................................................. 1

Chapter 2.

Quantitative imaging of the skeletal system for 3D reconstruction (Background) ............................................................................................ 7

2.1

Introduction ................................................................................................... 7

2.2

Computed tomography (CT) .......................................................................... 8

2.3

2.2.1

Basic principles of CT ........................................................................ 8

2.2.2

Radiation exposure during CT imaging ............................................... 9

Magnetic resonance imaging (MRI) .............................................................10 2.3.1

Basic principles of MRI .....................................................................10

2.3.2

How tissue contrast is determined ......................................................12

2.3.3

Selection of slice position and thickness ............................................13

2.3.4

Pulse sequences .................................................................................14

2.3.5

MRI safety.........................................................................................14

2.3.6

Signal to noise ratio of an MRI system...............................................14

2.3.7

Artefacts of MRI ................................................................................15

2.3.7.1

Motion artefacts ..........................................................................16

2.3.7.2

Magnetic susceptibility difference artefact ..................................16

2.3.7.3

Chemical shift ............................................................................17 VII

Contents

2.3.8

MRI for imaging of the skeletal system ............................................. 17

2.3.9

Advantages and current limitations of MRI ....................................... 18

2.4

2.3.9.1

Longer scanning times of MRI ................................................... 18

2.3.9.2

Poor contrast in certain anatomical regions ................................. 18

2.3.9.3

Non-uniformity of the external magnetic field ............................ 19

2.3.9.4

Limited accessibility .................................................................. 19

Summary ..................................................................................................... 20

Chapter 3.

Image processing and surface reconstruction ........................................ 21

3.1

Introduction ................................................................................................. 21

3.2

Acquisition of data for 3D modelling of bones ............................................. 22 3.2.1

Effect of in plane resolution and slice thickness on accuracy of reconstructed 3D models ................................................................... 23

3.3

Image segmentation ..................................................................................... 24 3.3.1

Manual segmentation ........................................................................ 25

3.3.2

Intensity thresholding ........................................................................ 25

3.3.2.1

Selecting an appropriate threshold level...................................... 26

3.3.2.2

Multilevel thresholding .............................................................. 26

3.3.3

Edge detection ................................................................................... 28

3.3.4

Region growing ................................................................................. 28

3.3.5

Sophisticated segmentation methods.................................................. 29

3.4

Surface generation ....................................................................................... 29

3.5

Registration (aligning) and comparison of surfaces ...................................... 30

3.6

A reference standard for validating 3D models of bones ............................... 30

3.7

Aims of the study ......................................................................................... 32

3.8

Methods ....................................................................................................... 32

VIII

3.8.1

Samples ............................................................................................. 32

3.8.2

Image segmentation........................................................................... 32

Contents

3.8.3

Reference model for validation of the outer 3D models ......................33

3.8.3.1

Removal of the soft tissues from long bones ...............................33

3.8.3.2

Scanning of the bone‟s outer surface using the contact scanner ...34

3.8.3.3

Reconstruction of the 3D model from scanned surfaces ..............37

3.8.4

Reference model for validation of the medullary canal .......................39

3.8.5

Basic 3D modelling techniques using Rapidform 2006 ......................41

3.9

3.8.5.1

Registration of 3D surfaces using Rapidform 2006 .....................41

3.8.5.2

Comparison of the aligned 3D models ........................................43

3.8.5.3

Dividing the 3D models of bones into anatomical regions ...........44

Results .........................................................................................................44

3.10 Summary, discussion and conclusion ............................................................45 3.11 Paper 1: Effect of CT image segmentation methods on the accuracy of long bone 3D reconstructions (published) .............................................................48 Chapter 4.

Application of 3D modelling techniques for orthopaedic implant design and validation ..........................................................................................57

4.1

Introduction ..................................................................................................57

4.2

3D models for implant design and validation ................................................58

4.3

Aims of the study .........................................................................................59

4.4

Methods .......................................................................................................59

4.5

Results .........................................................................................................59

4.6

Summary, discussion and conclusion ............................................................60

4.7

Paper 2: Quantitative fit assessment of tibial nail designs using 3D computer modelling (published) ...................................................................................61

Chapter 5.

Magnetic resonance imaging for 3D reconstruction of long bones ........67

5.1

Introduction ..................................................................................................67

5.2

Imaging of skeletal system with MRI ...........................................................68

5.3

Aims of the study .........................................................................................71

5.4

Methods .......................................................................................................71 IX

Contents

5.5

Results ......................................................................................................... 72

5.6

Summary, discussion and conclusion ........................................................... 72

5.7

Paper 3: Quantification of the accuracy of MRI generated 3D models of long bones compared to CT generated 3D models (in press) ................................ 74

Chapter 6.

Higher field strength MRI scanning of long bones for generation of 3D models ...................................................................................................... 83

6.1

Introduction ................................................................................................. 83

6.2

Theoretical consideration of increased SNR at 3T ........................................ 84

6.3

3T MRI for musculoskeletal system imaging ............................................... 84 6.3.1

Spin relaxation times and flip angle ................................................... 85

6.3.2

Fat suppression .................................................................................. 86

6.3.3

Magnetic susceptibility at 3T MRI..................................................... 87

6.3.4

Chemical shift at 3T .......................................................................... 87

6.3.5

MRI safety at 3T ............................................................................... 88

6.4

Aims of the study ......................................................................................... 88

6.5

Methods ....................................................................................................... 88 6.5.1

Samples ............................................................................................. 88

6.5.2

Measuring the quality of MR images ................................................. 88

6.5.3

Quantification of spin relaxation times .............................................. 90

6.5.4

Comparison of 1.5T and 3T imaging of musculoskeletal system ........ 91

6.6

Results ......................................................................................................... 93

6.7

Summary, discussion and conclusion ........................................................... 93

6.8

Paper 4: 3T MRI improves bone-soft tissue image contrast compared with 1.5T MRI (Submitted – under review).......................................................... 96

Chapter 7.

Step artefact caused by Magnetic Resonance Imaging of long bone ... 121

7.1

Introduction ............................................................................................... 121

7.2

Motion artefact of MRI .............................................................................. 122

7.3

Aims of the study ....................................................................................... 123

X

Contents

7.4

Methods ..................................................................................................... 123

7.5

Results ....................................................................................................... 124

7.6

Summary, discussion and conclusion .......................................................... 124

7.7

Paper 5: Correction of step artefact associated with MRI scanning of long bones (Submitted – under review) .............................................................. 126

Chapter 8.

Summary, conclusion and future directions.........................................145

8.1

Summary and conclusion ............................................................................145

8.2

Future directions......................................................................................... 148

Appendix 1 Ethical approval for the study in Chapter 6.........................................151 Appendix 2 Participant information and Consent form used in Chapter 6 ............154 Appendix 3 Animal tissue use notification ............................................................... 157 References

................................................................................................................ 159

XI

List of figures

List of figures Figure 2.1: Arrangement of the x-ray source, detector and the object in a CT scanner ...... 9 Figure 2.2: A spin possesses a tiny magnetic field aligned with the axis of rotation .........11 Figure 2.3: Spins aligned with the external magnetic field B0 ..........................................11 Figure 2.4: An MRI image of the coronal section of the proximal femur .........................19 Figure 2.5: The uniform regions of the external magnetic field of a MRI scanner (The uniform region is shaded) .....................................................................................19 Figure 3.1: Average intensity values of the outer bone contours as detected by the Canny filter for each axial CT image ...............................................................................27 Figure 3.2: The process of removing soft tissues from the sheep femur before scanning with the contact mechanical scanner: a - gross dissection with the scalpel, ............34 Figure 3.3: Scanning of the bone's outer surface of the diaphyseal region using the MDX 20 contact scanner (The bone is positioned on the stage using glue tags)...............35 Figure 3.4: Bone is cut in three parts in order to scan the articular surfaces which cannot be reached by the scanner on the intact bone .........................................................36 Figure 3.5: Positioning of the proximal articular segment of the femur in order to scan the articular surface ....................................................................................................37 Figure 3.6: The reconstructed model before the scanning of articular surfaces (This model was used as a guide to scan the articular regions) ..................................................37 Figure 3.7: Scanned surface with unusable data...............................................................38 Figure 3.8: The surface after removing the unusable data ................................................38 Figure 3.9: Two adjacent surfaces are fine registered ......................................................39 Figure 3.10: The final 3D model reconstructed by merging the surfaces ..........................39 Figure 3.11: a - The original microCT image (a cross section from the diaphysis); and b the image after applying a 20 × 20 median filter ...................................................40 XIII

List of figures

Figure 3.12: The initial aligning of the CT based 3D model to the reference model using Trackball prior to the application of fine registration function .............................. 42 Figure 3.13: A CT based model (red) is aligned to the reference model (blue) in Rapidform 2006 using the fine registration function ............................................. 42 Figure 3.14: Comparison of the aligned CT model to the reference model in Rapidform 2006 ..................................................................................................................... 43 Figure 3.15: Five anatomical regions used for the comparison: 1 - femoral head, 2 proximal region, 3 - diaphysis, 4 - distal region, 5 - distal articular region ............ 44 Figure 3.16: Reference planes and curves used for the splitting of the model into five anatomical regions ............................................................................................... 44 Figure 5.1: Cross sections of CT (left) and MRI (right) from the same anatomical location of a sample ........................................................................................................... 69 Figure 6.1: Positioning of the volunteer in the MRI scanner and the position of the matrix coils that cover the lower limbs and the pelvis ...................................................... 92 Figure 6.2: Positioning of the field of view (FOV) on volunteer‟s leg ............................. 93 Figure 7.1: The step artefact caused by volunteer moving the leg between two successive scanning stages................................................................................................... 121 Figure 7.2: MRI scanning of human lower limb with five scanning segments to scan the complete limb .................................................................................................... 123

XIV

List of tables

List of tables Table 3.1 Specifications of the MDX 20 contact 3D scanner ...........................................35 Table 3.2 Scanner parameters used for microCT scanning ...............................................40 Table 6.1 TR and TE values used for the MRI scanning at 1.5T and 3T ...........................90 Table 6.2 Different flip angles used for scanning .............................................................90 Table 6.3 The protocols used for MRI scanning ...............................................................92

XV

Publications, presentations and awards

Publications, presentations and awards Journal Publications 1. Rathnayaka K, Schuetz MA, Sahama T & Schmutz B. Correction of step artefact associated with MRI scanning of long bones, Submitted to Medical Engineering and Physics. 2. Rathnayaka K, Coulthard A, Momot K, Volp A, Sahama T, Schuetz MA & Schmutz B. 3T MRI improves bone-soft tissue image contrast compared with 1.5T MRI, submitted to Magnetic Resonance Imaging. 3. Rathnayaka K, Momot K I, Noser H, Volp A, Schuetz M, Sahama T & Schmutz B. Quantification of the accuracy of MRI generated 3D models of long bones compared to CT generated 3D models. Medical Engineering & Physics. 2011, in press, DOI:10.1016/j.medengphy.2011.07.027. 4. Rathnayaka K, Schmutz B, Sahama T and Schuetz M A. Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions Medical Engineering & Physics. 2011, 33(2): 226-233. 5.

Schmutz B, Rathnayaka K, Wullschleger ME, Meek J, Schuetz MA. Quantitative fit assessment of tibial nail designs using 3D computer modeling. Injury. 2010; 41(2): 216-219.

Conference presentations1 1. Rathnayaka K, Cowin G, Schuetz MA, Sahama T, Schmutz B. Correction of the step artefact in 3D bone models caused by the random movement of the lower limb during MRI. 17th Annual Scientific Meeting, Australian & New Zealand Orthopaedic Research Society. Brisbane, Australia, 1-2 September, 2011. (Oral presentation) 2. Rathnayaka K, Coulthard A, Momot K, Volp A, Sahama T, Schuetz M, Schmutz B. Improved image contrast of the bone-muscle interface with 3T MRI compared to 1.5T MRI. 6th World Congress on Biomechanics. Singapore, 1-7 August, 2010. (Poster presentation) 3. Schmutz B, Rathnayaka K, Wullschleger M, Meek J, Schuetz M. Quantitative fit assessment of tibial nail designs using 3 D computer modeling. German Society for Orthopaedic and Trauma Surgery. Berlin, Germany, 21-24 October, 2009. (Oral Presentation) 1

The conference abstracts have not been included in the thesis as the contents of them are covered by the journal articles

XVI

Publications, presentations and awards

4. Rathnayaka K, Sahama T, Schuetz MA, Schmutz B. Validation of 3D models of the outer and inner surfaces of an ovine femur. 15th Annual Scientific Meeting, Australian & New Zealand Orthopaedic Research Society. Adelaide, Australia, 9-10 October, 2009. (Oral presentation) 5. Rathnayaka K, Momot K, Volp A, Noser H, Sahama T, Schuetz MA, Schmutz B. Quantification of the accuracy of MRI generated 3D models of long bones. 4th Asian Pacific conference of biomechanics. University of Canterbury, Christchurch, New Zealand, 14–17 April, 2009. (Oral presentation) 6. Rathnayaka K, Schmutz B, Sahama T, Schuetz MA. Effects of image segmentation methods on the accuracy of long bone 3D reconstructions. 14th Annual Scientific Meeting, Australian & New Zealand Orthopaedic Research Society. Brisbane, Australia, 17-18 November, 2008. (Poster presentation)

7. Mohd Radizi S, Rathnayaka K, Pratap J, Mishra S, Schuetz MA, Schmutz B, The effects of CT convolution kernels on the geometry of 3D bone models. 14th Annual Scientific Meeting, Australian & New Zealand Orthopaedic Research Society. Brisbane, Australia, 17-18 November, 2008. (Poster presentation)

Awards and Scholarships 1. Outstanding HDR student of the month, Faculty of Built Environment and Engineering, Queensland University of Technology, December 2010. 2. Student travel grant awarded by 6th World congress on Biomechanics. Singapore, 1-7 August, 2010. 3. Joint winner of the Wilhelm-Roux-Preis 2009, at Annual conference of the German Society for Orthopaedic and Trauma Surgery. Berlin, Germany, 2124 October, 2009. 4. Student Travel grant awarded by 15th Annual Scientific Meeting of Australian & New Zealand Orthopaedic Research Society. Adelaide, Australia, 9-10 October, 2009. 5. Runner-up for best poster presentation, IHBI inspires postgraduate student conference. Gold Coast, Australia, 2-4 December, 2008. 6. QUT, Faculty of Built Environment and Engineering living allowance PhD scholarship 2008-2011. XVII

Authorship

Authorship I declare that the work contained in this thesis has not been previously submitted to meet the requirements for an award at this or any other higher education institution. To the best of my knowledge and belief, the thesis contains no materials previously published or written by another person except where due reference is made in the text. …………………………… Kanchana Rathnayaka

Date:……………….

Rathnayaka Mudiyanselage

XIX

Acknowledgement

Acknowledgement Firstly, I offer my sincere thanks to my supervisors: Dr Beat Schmutz for the invaluable support, guidance and advice given throughout the PhD and for helping to establish my directions; Prof. Michael Schuetz, my principal supervisor, for encouragement and guidance given; and Dr Tony Sahama for introducing me to the trauma research group at Queensland University of Technology and for the support given throughout the PhD study. I offer my special thanks also to: Dr Konstantin Momot for helping me by sharing his knowledge of MRI physics and by reading manuscripts, especially during the second and fourth parts of the research project; Prof. Alan Coulthard for collaborating with me for the fourth part of the project; Dr Gary Cowin for helping me with MRI scanning for the third part of the project; Mr Andrew Volp and Mr Russell Porter at Princes Alexandra Hospital; Mr Raymond Buckley at Royal Brisbane and Women‟s Hospital for MRI scanning of the samples and volunteers of the study; Mr Jit Pratap at Princes Alexandra Hospital; and Ms Margaret Day at University of Queensland for CT scanning of samples. I must offer my sincere gratitude to all those who volunteered as subjects for the study and spent their valuable time on my project, and to the National Imaging Facility for providing me with 100% subsidised access to the 3T MRI scanner. Thanks to all the researchers who donated ovine limbs from their studies and who helped me to obtain them at the Medical Engineering Research Facility (MERF). I would also like to thank the High Performance Computer (HPC) Unit and its personnel at QUT for their help with the 3D modelling and use of the super computers. Thanks to all the members of the trauma research group and all the friends who helped me with various aspects of this research, especially with feedback on writing and presentations. Finally, the laboratory and directorate staff at IHBI and MERF also provided kind help during this project and I offer them my gratitude.

XXI

Abbreviations

Abbreviations 13

C N 1 H 31 P 3D 3T B0 BW CAS CNR CT FA FOV H2O HU ICP M0 MHz MR MRI Mt Mz NAV NMR NPA NPE PMMA RF ROI SD SNR SNRGER SNRSE T1 T2 TE TMS TR V 15

XXII

Carbon Nitrogen Hydrogen Phosphorus Three dimensional Three tesla Main magnetic field Bandwidth Computer assisted surgery Contrast to noise ratio Computed tomography Flip angle Field of view Water Hounsfield units Iterative closest point Net magnetisation vector Mega Hertz Magnetic resonance Magnetic resonance imaging Transverse component of net magnetisation vector Longitudinal component of net magnetisation vector Number of signal averages Nuclear magnetic resonance Number of acquired partitions Number of acquired phase encodes Poly-methyl methacrylate (Dental acrylic) Radiofrequency Region of interest Standard deviation Signal to noise ratio Signal to noise ratio for gradient echo sequence Signal to noise ratio for spin echo sequence Longitudinal relaxation time Transverse relaxation time Echo time Tetramethylsilane Repetition time Voxel volume

Chapter 1: Introduction

Chapter 1 Introduction The introduction of x-ray computed tomographic (CT) scanning and magnetic resonance imaging (MRI) in the 1970s allowed medical personnel and researchers to visualise the internal anatomical structures of the human body in three dimensions. This allowed clinicians and researchers to reconstruct anatomical structures as computer based three dimensional (3D) models and perform various experiments that cannot be performed on living subjects. Thus, accurate reconstruction of 3D models of anatomical structures from CT and MRI became a major research interest. Even though the main mode of imaging bones is CT, the involvement of ionising radiation leads clinicians and researchers to avoid CT whenever possible. Thus, a trend towards the frequent use of MRI is developing among these groups, not only due to the non-involvement of ionising radiation in MRI, but also due to its ability to provide better quality images of soft tissue. Reconstruction of a three dimensional computer model of an anatomical structure using either CT or MRI imaging methods involves a number of complex processes: data acquisition; segmentation of the region of interest (ROI) and surface generation from the segmented volume. Each of these processes plays a crucial role in determining the geometric accuracy of the reconstructed 3D model. Since the geometric accuracy of 3D models is of high importance for most of their applications (e. g. implant design and simulation of surgery), these processes in reconstructing 3D models have drawn major attention from researchers [1-3]. While all steps play a crucial role in determining the accuracy of 3D models, image segmentation is one of the steps which has a higher human involvement and is thus vulnerable to errors. Even though existing sophisticated segmentation methods are capable of minimising the human intervention, most of these methods involve complex programming and mathematics which many of the researchers are not trained to perform [2, 4-7]. 1

Chapter 1: Introduction

Furthermore, these algorithms are designed to perform segmentation in a specific anatomical region and, therefore, are not easily extended to the segmentation of a different region due to their complex nature. Thus, a simple but accurate method for medical image segmentation is a necessity. Reconstruction of a 3D model of a small bone (phalanges or metatarsal bones) is relatively easy when compared to the reconstruction of a 3D model of a long bone that has a complex geometry. Thus, most of the studies that investigated segmentation methods have utilised small bones. Nevertheless, 3D reconstruction of long bones is important as most of the fracture fixation plates and intramedullary nails are used for fixation of long bones. When 3D models of long bones are reconstructed, the diaphyseal as well as the distal and proximal regions are equally important. Most of the fracture fixation plates and intramedullary nails extend to the proximal and distal regions (e.g. expert tibia nail used in chapter 4). The intramedullary nail insertion point is usually in the proximal or distal region, thus, accuracy of these regions are important to determine the entry point of the nail. Furthermore, design of implants such as joint replacements needs highly accurate 3D models of the proximal and distal articular regions. Therefore, the research projects contained in the thesis will focus on all anatomical regions of long bones. The decision to use either CT or MRI is mainly determined by the anatomical structures being scanned. While CT visualises the bone tissue with better contrast, MRI visualises soft tissues with better contrast as its main source of signal is hydrogen nuclei which are abundant in soft tissues. The radiation exposure of CT limits its utilisation to clinical cases and cadaver specimens. As most of the available cadavers are more than 60 years old, the data acquisition is also limited to this age group. However, approximately 51% of land transport trauma patients (or 11.4% of total injury hospitalisations) in Australia during the 2006-2007 period were under 30 years of age. Furthermore, the study conducted by Noble et al 1995 [8] shows that the femoral isthmus expands in old female population compared to the young population. The study also showed that the medullary canal expands and the cortex becomes thinner in old females and the CCD angle (femoral neck-shaft angle) change with the age. These changes will impact the anatomical fitting of plates and intramedullary nails designed using 3D models reconstructed from old bones. In addition, osteophytes in old bones can significantly affect the anatomical fitting of 2

Chapter 1: Introduction

fracture fixation plates especially in ends of the bones. Thus, the acquisition of bone data from this age group to inform the design of anatomically shaped fracture fixation implants (plates and nails) for its trauma patients is of utmost importance [9]. As MRI does not utilise ionising radiation, it is a potential alternative to CT for acquiring bone data of volunteers from younger age groups. Even though MRI visualises soft tissues with a high contrast, due to the extremely short transverse relaxation times, bones generally do not generate a signal in MRI [10-12]. However, using the signal generated by the soft tissues, bone geometry can be delineated from the surrounding soft tissues and this has been demonstrated in the literature [1, 13-17]. MRI has been used for the scanning of bones mainly in the case of diagnosing metastatic disease, as MRI visualises metastasis with better quality [18]. The use of MRI for 3D reconstruction of bones has been reported in computer assisted surgery (CAS) and in foot bone motion quantification where the 3D models of vertebrae and tarsal bones have been reconstructed [19-21]. Most of these studies have used MRI for small bones with relatively simple geometry, and a proper validation of the models has not been performed. Lee et al. used MRI to generate a 3D model of a porcine femur; however, the model has not been validated using an accurate validation standard [1]. Therefore, before using MRI for 3D reconstruction of long bones, a quantitative validation with an accurate reference standard is necessary. Some of the current limitations of the MRI scanning of long bones are long scanning times and the difficulty of segmenting certain anatomical regions, conferred by poor contrast between those anatomical regions and surrounding soft tissues. Since the signal to noise ratio (SNR) of an MRI system is approximately directly proportional to the main magnetic field of the scanner, higher field strength (3T) scanners promise to offer an improved signal which can be converted to faster scanning times or better image quality compared to the currently available (1.5T) scanners [22, 23]. The improved image quality of 3T scanners has been demonstrated in a few studies for computer assisted surgery and kinematic analysis of foot bone motion [24, 25]. However, the contrast at the bone muscle interface, which is more important for segmentation of bones, has not been quantified and compared in those studies. 3

Chapter 1: Introduction

Furthermore, different contrast levels which occur in different anatomical regions of a long bone need to be studied in detail to see the improvement in contrast at those regions. MR imaging of anatomical structures is challenged by various artefacts. Among them, the motion artefacts due to random movements are of main concern in MRI imaging of long bones due to their effects on the geometric accuracy of 3D models reconstructed. In addition to the long scanning time of MRI, the non-uniformity of the main magnetic field limits the effective scanning length, resulting in a long bone being scanned in several stages. One of the adverse effects of this, which has been observed in an initial study conducted by the supervisory team, is the displacement artefact caused by the volunteer moving the leg between two scanning stages [26]. Thus, there is a step that can be seen on the final 3D model generated from such a data set. This artefact may not be critical for clinical use of the images; however, when the precise measurements are performed for implant design, these artefacts can have a major effect on their accuracy. Therefore, minimisation or correction of these artefacts can improve the accuracy of implants designed using those models. This thesis presents the studies carried out to investigate: a simple and accurate method for medical image segmentation; the feasibility of MRI as an alternative to CT for scanning of long bones; the usability of higher field strength MRI to overcome some of the problems with low field strength scanners; and the correction of the step artefact that occurred from MRI scanning of long bones. Chapter 2 provides the basic physics involved in CT and MRI, while Chapter 3 provides the background of image segmentation, 3D reconstruction and the investigation carried out to develop and validate a simple and accurate image segmentation method. Chapter 4 presents the application of 3D modelling techniques in implant validation, utilising 3D models of long bones for fit quantification of two anatomically shaped intramedullary nails. Chapter 5 presents the investigation carried out to formally validate the MRI based 3D models of long bones against the CT based models. Chapter 6 provides the details of the quantitative comparison between 1.5T MRI and 3T MRI. Chapter 7 presents the correction of the step artefact that occurred due to the random movements of the volunteer during MRI scanning. Chapter 8 presents a summary, discussion and future directions of the thesis.

4

Chapter 1: Introduction

The aims of the study in brief are as follows: Investigation of the accuracy of multilevel intensity thresholding and Canny edge detection for segmentation of CT images Quantification of the accuracy of 3D models based on MRI compared to the 3D models based on CT Quantitative comparison of the image quality at 1.5T MRI to 3T MRI Correction of the step artefact that occurs due to the random movement of the lower limb during MR imaging

5

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

Chapter 2 Quantitative imaging of the skeletal system for 3D reconstruction (Background) 2.1 Introduction A number of methods are available for the imaging of various anatomical structures of the human body, such as: plain x-ray, computed tomography (CT), Dual energy xray absorptiometry (DEXA), magnetic resonance imaging (MRI) and ultrasound (US). Even though quantitative imaging of the skeletal system is possible with most of the above scanning methods, accurate spatially-resolved information of the anatomical structures can only be acquired using CT or MRI. Thus, CT and MR imaging methods have taken an integral part in research and in clinical applications where the 3D reconstruction of the anatomical structures is required. The most commonly used imaging technique for quantitative imaging of the skeletal system is CT; however, MRI has also been reported as a potential imaging technique for this purpose. CT has become the gold standard of imaging the skeletal system for 3D reconstruction because CT produces images with better contrast at the bone–soft tissue interface. CT images can also be acquired within a very short period of time, thus, essentially avoiding the motion artefacts caused by moving body parts or tissues. While CT involves ionising radiation that prevents its use on healthy human volunteers for research purposes, it can be used for in vitro research studies for scanning of bones. In the present study, CT will be used to validate two image segmentation methods and for validation of MRI based models of ovine femora. Section 2.2 of this chapter provides the basic principles of CT and discusses its advantages and disadvantages. 7

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

MRI utilises the principles of nuclear magnetic resonance (NMR) of hydrogen nuclei to generate a signal from the tissue. Even though the bone tissues do not generate a significantly large signal, by utilising the signal generated from the surrounding soft tissues, MRI can be used in the 3D reconstruction of bones. Thus, MRI can potentially be used to image healthy human volunteers for research purposes, without having to expose them to the ionising radiation of CT. Section 2.3 discusses the basic principles, advantages and current problems of MRI in detail. Since MRI has been utilised for most of the studies presented in this thesis, it will be discussed in more detail than CT.

2.2 Computed tomography (CT) Computed tomography (CT) was the first method of imaging anatomical structures inside the body without having the problem of the superimposition of anatomical structures that was a major drawback of plain X-ray images. Since its introduction to clinical use in 1970 [27], CT has become the most commonly used imaging technique in the clinical setting. It has also become the standard practice for imaging of trauma patients for accurate diagnosis of bone fractures in emergency situations [28, 29].

2.2.1 Basic principles of CT CT images are acquired by recording the object‟s attenuation of the radiation which is emitted from an x-ray source (x-ray tube). A CT image is reconstructed from a large number of projections of the object, taken around a single axis of rotation using an x-ray beam. Depending on its x-ray absorption properties, when the x-ray beam passes through the object, a projected image is generated on the detector (image sensor). These images are integrated using a computer based algorithm to produce axial image slices. The projections are obtained by rotating the detector and the x-ray source simultaneously around the object (Figure 2.1).

8

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

Figure 2.1: Arrangement of the x-ray source, detector and the object in a basic CT scanner (A large number of projections of the object will be obtained by rotating the source and the detector simultaneously around the object.) Early generation CT scanners imaged a patient slice by slice with specific slice spacing. Once an image slice is obtained, the table with the patient moves a set distance and the next slice is obtained. With the development of helical CT, continuous imaging is performed by moving the patient continuously through the gantry in combination with the continuous rotation of the x-ray source and detector system. The obtained data volume is later reconstructed to image slices with specific slice spacing. This also allows for the reconstruction of images in anatomical planes other than the traditional axial image slices. Modern spiral scanners with multiple rows of x-ray detectors (multi-slice scanners or multi-row scanners) can image a subject within a very short time period (a few seconds), thus almost eliminating motion artefacts. Due to the high accuracy obtained for the bone geometry, CT has become the gold standard for imaging of the bones for reconstructing 3D models, mainly for the development of implants and clinical applications. CT can also be used for measurement of relative tissue density and can be presented as Hounsfield Units (HU) for comparison with other or reference tissues.

2.2.2 Radiation exposure during CT imaging The use of diagnostic CT has increased dramatically over the last 20 years and it is the gold standard for bone imaging. However, CT uses a high dose of radiation and concerns have been raised regarding cumulative radiation exposure and associated lifetime risk as there is epidemiological evidence of a small risk of radiation associated cancer at doses comparable to a few CT scans [30-33]. For example, the radiation exposure of a standard thoracic CT is equivalent to 400 standard chest x-ray 9

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

radiographs (8mSv), and that of a pelvic CT is equivalent to 250 chest radiographs (250 mSv) [30, 34, 35]. According to a report by the Royal College of Radiologists in the UK, CT scans probably contribute almost half of the collective dose of radiation from all x-ray examinations [36]. This has become a major problem, as CT scanning of a healthy human volunteer for research purposes is ethically not justifiable due to this high radiation exposure. As a solution, protocols that use low radiation doses while maintaining a higher image quality are under investigation [37-39]. Some slice selection strategies (e.g. use of fewer slices for simple geometric shapes such as diaphyseal region) have also been investigated to reduce the radiation dose [38]. However, due to the fact that the radiation exposure of CT cannot be eliminated completely, some countries do not approve the scanning of volunteers with these protocols. Therefore, researchers are moving towards using an imaging technique such as MRI that does not utilise ionising radiation.

2.3 Magnetic resonance imaging (MRI) 2.3.1 Basic principles of MRI Magnetic resonance imaging (MRI) utilises the nuclear magnetic resonance (NMR) of 1H nuclei as the source of signal. There are a number of elements that demonstrate NMR capabilities, such as 1H,

13

C,

15

N,

31

P. Human tissue is largely composed of

water (H2O) and, thus, 1H is the most abundant NMR capable nuclei in the human tissue. Throughout this discussion, 1H nuclei are also referred to as „spins‟, as 1H nuclei have the quantum mechanical property termed „nuclear spin‟. If a single 1H nucleus is considered, it possesses a magnetic moment, which is a quantum mechanical property, parallel to its axis (Figure 2.2). In the absence of an external magnetic field, the axes of the spins are randomly aligned in a given tissue sample and the vector sum of the magnetisation is equal to zero (Figure 2.2).

10

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

Figure 2.2: A spin possesses a tiny magnetic field aligned with the axis of rotation (left); randomly aligned axes of spins in the absence of an external magnetic field (right) [40]2 To measure NMR of 1H nuclei (or any NMR capable nucleus), an external magnetic field (also referred to as „the main magnetic field‟ or „B0‟) is applied to the sample, thus making randomly aligned spins partially align with the externally applied magnetic field (in the opposite direction to B0) (Figure 2.3). Thus, the sample now possesses a net magnetisation vector (M0) parallel to B0. M0 can be split into its component vectors: Mz which is parallel to B0, and Mt which is perpendicular to B0. At rest, Mz = M0 and Mt = 0 (Figure 2.3).

Figure 2.3: Spins aligned with the external magnetic field B0 and M0 and its two components, Mz and Mt. 2

Adapted from: Brown, M.A. and Semelka, R.C. MRI Basic principles and applications, 4th ed. 2010, New Jersey: John Wiley & Sons

11

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

The frequency at which the spins precess under an external magnetic field is proportional to the strength of the external magnetic field and is expressed by the Larmor equation (Equation 2.1):

v0

B0 2

2.1

Where v0 is the Larmor frequency in megahertz (MHz), B0 is the external magnetic field strength in tesla (T) and

is a constant known as gyromagnetic ratio [40].

If an external radiofrequency (RF) wave with a frequency same as the Larmor frequency of the spins (~64 MHz at B0 = 1.5T) is applied to the sample, some of the spins shift from a low energy orientation to a high energy orientation. This moves M0 of the spins towards a direction perpendicular to B0 (if a 90º pulse is applied), generating a net transverse magnetisation (Mt), and leading Mz to decline. At this stage, the Larmor precession of the spins will induce a voltage in the receiving coil (RF coil) which is measured as the MR signal. The intensity of the signal generated in the receiver coil is proportional to the transverse magnetisation (Mt); therefore, the initial magnitude of the signal depends on the value of the Mz immediately prior to the RF pulse. When excited, the angle at which M0 is oriented relative to B0 is the flip angle (FA), which is one of the parameters that should be changed accordingly to get an optimal contrast. When the RF wave is shut off, Mz starts to recover, and the inverse of the rate constant of recovery is called the „longitudinal relaxation time‟ (T1). At the same time, Mt starts to decay and the exponential rate constant of decay is called „transverse relaxation time‟ (T2). Both T1 and T2 take different values for different tissue types [41].

2.3.2 How tissue contrast is determined In MRI, tissue contrast is related to the differences of rate of magnetisation decay. The three factors that determined the tissue contrast in the present work were T1, T2 and the proton density of the tissue. The differences between spin relaxation times and the proton density in different tissues serve as the basis for image contrast. The contrast can be manipulated by selecting different scan parameters, namely repetition time (TR) and echo time (TE). TR is the time period between two successive 12

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

acquisitions and TE is the time between delivery of the RF pulse and signal detection. T1 contrast can be selected by choosing short repetition times (TR) and such images are called „T1 weighted images‟ where the contrast is mainly determined by T1 of the particular tissue. T2 contrast can be modulated by changing echo time (TE) and the images of which the contrast is mainly determined by T2 are called „T2 weighted images‟. In both types of images, there is a contribution from T 1 and T2, however, the effect from one is minimised while the other is maximised.

The

contrast can also be determined by the proton density of the tissue and the images acquired this way are called „proton density weighted images‟.

2.3.3 Selection of slice position and thickness The slice position, slice thickness and the Phase and Read directions are determined by the respective gradient pulses and (in the case of the slice position) the RF frequency offset. When a magnetic field gradient is applied on top of the existing main field B0 in x, y, or z directions, the spins at different locations along the gradient experience slightly different magnetic fields. Thus, the spins at different locations along the gradient precess at different Larmor frequencies, which are given by the following equation (2.2):

vi

( B0 G ri )

2.2

Where vi is the frequency of the spin at position ri , G is the gradient vector representing the total gradient amplitude and the direction, B0 is the main magnetic field and

is the gyromagnetic ratio [40].

Three linear mutually perpendicular gradients are used: phase encoding gradient, readout gradient and slice selection gradient. The phase-encoding gradient encodes the locations of the nuclei in the direction of that gradient using the phase accumulated by the nuclei during the gradient pulse. The readout (or frequencyencoding) gradient encodes the locations of the nuclei in the direction of that gradient using the position-dependent precession frequency during acquisition of the echo. The receiver coils detect the entire spectrum of the different precession frequencies during the readout gradient, which ensures that the field of view in the Read direction covers the entire sample.

13

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

The slice selection gradient is used to achieve the localisation of the RF excitation to a region in the space. The RF pulse applied has two parts: a central frequency and a narrow bandwidth of frequencies (1-2 kHz). When such RF pulse is applied to the sample in the presence of the slice selection gradient, a narrow region of tissue achieves the resonance state. Thus the bandwidth of the applied RF wave determines the thickness of the image slice.

2.3.4 Pulse sequences The pulse sequence is a sequence of instructions to the hardware for switching the RF pulse and gradient pulses on and off and for sampling the signal, keeping a specific time period between each of them. This allows for the acquisition of data in the desired manner by manipulating the relevant parameters (TR, TE, and FA). Spin echo sequence and gradient echo sequence are two commonly used sequences for clinical imaging. The FLASH (Fast Low Angle Shot) sequence used for this study is based on the gradient echo sequence.

2.3.5 MRI safety MRI is relatively safe compared to CT; however, the RF power deposition in the conductive tissue results in heating of the tissue inside the body. To prevent hazards from the heat, the specific absorption rates (SAR) of energy dissipation are monitored using hardware level or software level monitors [42]. There are no known direct biological hazards to patients from exposure to strong magnetic fields. However, there is a high risk of the strong magnetic field of the scanner affecting metallic implants and cardiac pacemakers. Thus, MRI is contraindicated for patients with cardiac pacemakers, metallic debris in eyes or other ferromagnetic materials in the body. Patients with implants that do not have a risk of detaching, or which do not contain ferromagnetic materials (e.g. hip replacements, stents made of nickeltitanium alloy) can be safely scanned with MRI [27].

2.3.6 Signal to noise ratio of an MRI system Signal to noise ratio (SNR) is an important measure that can be used to quantify the quality of a MRI system (Equation 2.3). In the case of conducting tissues, the intrinsic SNR of a MRI system is approximately proportional to the strength of the external magnetic field and the volume of tissue being scanned, and depends on tissue parameters (e.g. T1 & T2). The following equations (2.4 & 2.5) show the 14

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

relationship between SNR and other parameters of a MRI system for a spin echo sequence and gradient echo sequence [43]:

SNR

SNRSE

SNRGER

Signal Noise

B0V

B0V

2.3

N PE N PA N AV (1 e BW

TR T1

)e

TE T2

N PE N PA N AV sin (1 e TR T1 ) e BW (1 e TR T1 ) cos

2.4

TE T2

2.5

Where SNRSE = signal to noise ratio for spin echo sequence, SNRGER = signal to noise ratio for a gradient echo sequence (FLASH), B0 = external magnetic field, V = voxel volume, NPE = number of acquired phase encode lines, NPA = number of acquired partitions, NAV = number of signals averaged, BW = receiver bandwidth per pixel, TR = repetition time, T1 = longitudinal relaxation time, TE = echo time, T2 = transverse relaxation time and θ = flip angle. In both equations, the term under the square root is the total time for acquiring data. Therefore, intrinsic SNR is directly proportional to the strength of the external magnetic field, the voxel volume, the square root of total sampling time and contrast related parameters. Thus, from the above relationship, it is clear that the external magnetic field, voxel size, number of averages, flip angle, T1, T2, TR and TE all have an influence on SNR of a MRI system. In addition, sensitivity to magnetic susceptibility and chemical shift difference between fat and water also influence the SNR of a MRI system.

2.3.7 Artefacts of MRI When the pixels in the MR image do not represent the actual anatomical structure being scanned, this region of the image is referred to as an „artefact‟. These artefacts appear among the general structures as signals that do not correspond to the actual tissue at the location. They may or may not be easily recognised from the normal anatomy, particularly if they are low in intensity.

15

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

2.3.7.1 Motion artefacts Motion (also referred to as movement) artefacts occur as a result of movement of the tissue (heart, lung) or parts of the body (limbs) which are being scanned during the data acquisition. Motion artefacts can either be due to periodic movements (e.g. blood flow, respiration and heart beat) or random movements which mainly occur due to the person‟s inability to keep the body parts still during the long scanning time. These movements result in misregistration of pixels along the phase-encoding direction [40, 44]. The artefact occurs by tissue that is excited at one location producing signals that are mapped to a different location during detection [40]. The nature and the extent of the artefact depend on the extent of movement and the protocol used for scanning. The most common motion artefact caused by periodic movements is due to blood flow in the vessels of the tissue being scanned [40]. If the blood flow is in a direction perpendicular to the slice plane, the artefact is localised to the vessel diameter. If the flow is along the slice plane, a more diffuse artefact is seen. Motion artefacts from random movements occur due to muscle contraction from nerve excitations. They can also occur as a result of the volunteer or patient randomly moving the body part being scanned due to the longer scanning times (e.g. keeping a lower limb still for 65 minutes is nearly impossible). Since the complete lower limb has to be scanned in several segments, volunteers tend to move the leg between segments and this causes a step in the final image stack. Motion artefacts that occur due to periodic movements such as breathing movements can be minimised by using specially designed protocols which synchronise the data acquisition with the breathing movements, or by post processing techniques. Elimination of the artefacts occurring due to random movements is, however, more difficult to achieve through such methods. 2.3.7.2 Magnetic susceptibility difference artefact Magnetic susceptibility ( ) is the response of a substance to the applied magnetic field. There are three levels of responses that have been described: diamagnetic, paramagnetic and ferromagnetic. The diamagnetic response arises from the electrons surrounding the nuclei, while the paramagnetic response arises from molecules that have unpaired electrons. Both these responses are relatively weak responses and 16

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

materials with such responses are safe to be used in MRI. However, the ferromagnetic response is found in certain ferrous metals and the magnetic susceptibility due to this response is very large. The relationship between magnetic susceptibility, external magnetic field and net magnetisation vector is expressed as the equation below:

M0

B0

Where M0 = net magnetisation vector,

2.6

= magnetic susceptibility and B0 = external

magnetic field [40]. The artefact is generated due to the different magnetic susceptibility of two adjacent tissue types. Cortical bone has a low magnetic susceptibility, while soft tissues have larger magnetic susceptibility. Thus, at the interface between soft tissue and bone, a considerable change in the local magnetic field present causes a significant signal loss. 2.3.7.3 Chemical shift Chemical shift is the difference in precessional frequency conferred by the magnetic shielding effect of the electron clouds that surround protons within tissues, relative to that of a standard reference compound (in the case of protons tetramethylsilane ( TMS)). Basically, in MRI there are two sources of 1H nuclei, water and fat. Water has two H atoms bonded to one oxygen atom, while fat has many H atoms bonded to a long-chain carbon framework. Due to this difference, protons from water have a different local magnetic field than protons from fat which is called „magnetic shielding‟. This magnetic shielding effect causes the protons from two sources to precess at different frequencies. This, in turn, causes fat and water protons from the same tissue location map to different positions in the reconstructed image. The difference of precessional frequency between water and fat at 1.5T is approximately 220Hz.

2.3.8 MRI for imaging of the skeletal system MRI is designed to scan soft tissues utilizing 1H nuclei as the source of signal and, thus, is not routinely used for imaging of bones. However, by using the signal generated from the surrounding soft tissue, bone outer geometry can be quantified from MRI images. This will be discussed in more detail in Chapter 4. 17

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

2.3.9 Advantages and current limitations of MRI Absence of ionising radiation is an important advantage of MRI over CT, as this allows researchers to scan healthy human volunteers without exposing them to a high dose of ionising radiation. However, MRI has some limitations compared to CT when used for scanning of long bones, such as longer scanning times, poor contrast in certain anatomical regions, non-uniformity of the magnetic field, limited availability, and higher cost per scan. 2.3.9.1 Longer scanning times of MRI Longer scanning time is the most important limitation of MRI when it is used for scanning of clinical cases as well as for research. As an example, in this study, scanning of a human lower limb with a modern 64 slice helical CT scanner takes less than ten seconds of scanning time, while an MRI scanner takes more than one hour for the same scan. This longer scanning time of MRI makes the images of moving (breathing) body parts vulnerable to motion artefacts. 2.3.9.2 Poor contrast in certain anatomical regions The next important limitation of MRI is the poor contrast of MRI images in certain anatomical regions of the bone (Figure 2.4). In the human body or other mammalians (sheep), the diaphyseal region of long bones is covered mostly with muscles. However, the distal and proximal regions of the bone, on the other hand, are mostly covered with ligaments, joint fluid, joint capsule and cartilage. These different soft tissue types have different MRI properties and, depending on the chosen scanning parameters, some generate poor or no signal, thus making them indistinguishable from cortical bones (e.g. ligaments, cartilage). Thus, the demarcation between such soft tissues and the cortical bone cannot be clearly defined and a complete 3D model is generated by making an educated guess or by interpolating the available data. This educated guessing or interpolation of the regions introduces errors to the 3D models.

18

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

Figure 2.4: Left: an MRI image of the coronal section of the proximal femur (showing that the shaft region has a good contrast between cortical bone and the muscles, while the regions indicated by arrows are not clearly defined), right: the corresponding CT image (showing the well defined boundary of cortical bone) 2.3.9.3 Non-uniformity of the external magnetic field The external magnetic field used by MRI scanners is not uniform throughout its length (Figure 2.5). Due to this non-uniformity of the magnetic field, the signal of the MRI images tends to distort towards the ends of the magnetic field, thus limiting the effective scanning length of the scanner to about 30 - 40 cm. Therefore, long samples (such as human lower limbs) have to be scanned in several stages; this involves moving the table to position different parts of the sample in the centre of the magnet.

Figure 2.5: The uniform regions of the external magnetic field of a MRI scanner (The uniform region is shaded) 2.3.9.4 Limited accessibility The accessibility of MRI scanners for research is mainly determined by the cost and the clinical work load of the scanner. The cost of a MRI scan is considerably higher than the cost of a CT scan. Due to the longer scanning times, MRI scanners in clinical use are heavily booked for scanning of patients. Few scanners are dedicated for research purposes. With the increased clinical use of 3T MR imaging, more 19

Chapter 2: Quantitative imaging of the skeletal system for 3D reconstruction (Background)

scanners will become available and the availability of scanners for research purposes will potentially be increased.

2.4 Summary The two available scanning methods for quantitative 3D imaging of long bones are CT and MRI. CT and MRI both provide accurate information for quantifying anatomical structures in a 3D environment. Both imaging methods have certain uses in clinical application, with CT mainly being used for bone imaging and MRI for soft tissue imaging. CT uses ionising radiation for its scanning and, therefore, its use in research is generally limited to scanning of cadaver specimens or clinical cases. While CT has a number of advantages such as a high contrast in bone–muscle interface and faster imaging times, it cannot be used for scanning of human volunteers. MRI utilises NMR of the 1H nuclei as the source of signal for imaging. Hence, the theoretical use of MRI is limited to imaging of soft tissues. However, MRI has the advantage of not using ionising radiation and is therefore well suited for scanning of healthy human volunteers for research purposes. MRI has some limitations such as very long scanning times, poor contrast in certain anatomical regions and shorter scanning length due to non-uniformity of the magnetic field. Limitations such as longer scanning times and poor contrast in certain anatomical regions can be overcome to some extent by using an external magnetic field with a higher strength, as later demonstrated by the study conducted in Chapter 6.

20

Chapter 3: Image processing and surface reconstruction

Chapter 3 Image processing and surface reconstruction 3.1 Introduction The reconstruction of 3D models of bones using CT imaging has become an interest of current medical engineering researchers, as 3D models of bones are increasingly being utilised for various practices of clinical medicine and medical research; for example, in the design of orthopaedic implants [45-50], in the computer aided planning of surgery [51-54], in fracture healing models [55, 56] and in finite element methods for fracture load analysis and bone strength analysis [57-61]. This interest is not only due to the wide utilisation of 3D models, but also because the generation of an accurate 3D model is a complex process. This process involves several stages where accuracy of the 3D model can be highly affected by various factors in each stage. The process of reconstructing a 3D model from a bone can be categorised in terms of data acquisition, image segmentation and surface generation [62]. Data acquisition is conducted using a tomographic imaging method such as CT or MRI. Image segmentation is the process of separating the identified ROI. The surface generation of the segmented volume is performed automatically using an algorithm, and is one step that determines the sub-voxel level accuracy of 3D models. While all steps play a crucial role in generating an accurate 3D model, it is the segmentation step that is most user-dependent and thus vulnerable to operator introduced inaccuracies. Thus, an accurate image segmentation method is of utmost importance for generating 3D models with correct geometric representation of the actual bones. Among the various segmentation methods available, intensity thresholding and edge detection are two simple image segmentation methods commonly used in medical 21

Chapter 3: Image processing and surface reconstruction

image segmentation. This study investigates intensity thresholding and Canny edge detection as simple but accurate segmentation methods for long bone image segmentation that can be used by the general research community who do not have a background in the complex programming and mathematics involved in segmentation. The next section discusses the relevant literature and the processes involved in reconstructing 3D models of bones from a CT or MRI data set. From Section 3.8 onwards the description will be focused on the 3D modelling methods used in this study. The image segmentation methods investigated and the surface generation and 3D model manipulation techniques described in this chapter are used in all the projects that are included in this thesis.

3.2 Acquisition of data for 3D modelling of bones In imaging, data acquisition is the process of obtaining a digital representation of the anatomical structures. While this can be achieved using various acquisition techniques, the acquisition of data from living subjects for 3D reconstruction of bones is done by using CT or MRI imaging, as other methods cannot generate 3D spatially resolved information of the anatomical structures. CT and MRI can also be used for acquisition of the image data from cadaver bone specimens, and CT is the gold standard for this process. The soft tissue free cadaver bones can be scanned with contact mechanical scanners or optical 3D scanners. In certain countries CT cannot be used to scan healthy humans for research due to the high amount of radiation involved in CT. The accuracy of the data acquired from bones depends on the type of imaging method, the accuracy of the hardware used and the set imaging parameters. Adequately calibrated hardware and optimally set scanning parameters are necessary for accurate acquisition of data from any anatomical structure. The calibration of the hardware has usually been conducted at the factory and the recalibration is conducted periodically by scanning phantoms. The scanning parameters vary with the imaging modality (e.g. MRI or CT) and can be adjusted depending on the structures to be visualised. For reconstruction of 3D models, the data should be acquired as spatially resolved information of the anatomical structures. Tomographic imaging techniques such as 22

Chapter 3: Image processing and surface reconstruction

CT and MRI are capable of obtaining such data of anatomical structures. Generally in CT or MRI, 3D volumetric data is presented as axial image slices with a certain user defined thickness. A single image slice is composed of a two dimensional array of elements called „pixels‟ and with the thickness added, these elements are called „voxels‟. The voxel basically represent the average signal or intensity of the tissues contained in it. The size of a voxel is determined by the field of view (FOV), size of the image matrix and the slice thickness.

3.2.1 Effect of in plane resolution and slice thickness on accuracy of reconstructed 3D models In order to obtain an accurate representation of the anatomical structures being scanned, a voxel should be sufficiently small in size. When the voxel size becomes larger, it contains the average signal/attenuation from larger tissue volume and more tissue types. Hence, larger voxel size (low resolution) results in higher inaccuracies in 3D models due to inadequate representation of the anatomical structures. This mainly affects the scanning of thin structures (e.g. distal and proximal regions of the cortex of a long bone) where the thickness is less than the voxel size. In the case of CT, this results in overestimation of the thickness of the structure and underestimation of its density [63, 64]. This produces 3D models that do not accurately represent the surface geometry affecting the implants generated using such models [65, 66]. In addition, the bone density properties acquired from such data adversely affect the accuracy of the FE models. This inaccurate representation of anatomical structures by pixels is called „the partial volume effect‟. The partial volume effect appears when one element (voxel) is filled by tissues with different attenuation properties for which the mean attenuation is calculated [67]. Appearance of this effect in the bone–muscle interface makes the separation of the bone a relatively difficult process requiring more robust segmentation methods. Generally, this effect can be minimised using a smaller voxel size [68]. However, the complete elimination of this effect is not possible. In addition, acquisition of the image data with smaller voxel sizes is done at the expense of imaging time in the case of MRI, or of exposing the subject to a high radiation dose in the case of CT.

23

Chapter 3: Image processing and surface reconstruction

The slice spacing or thickness also has an effect on the reconstructed 3D models, especially when relatively complex shapes have to be reconstructed (e.g. femoral head) compared to simple shapes (e.g. diaphysis of a long bone) [69-71]. This effect can be minimised using a smaller slice spacing or thickness, especially where the geometric shapes are complex [69]; again, however, this is at the expense of imaging time in MRI or radiation exposure in CT.

3.3 Image segmentation Image segmentation is the process of separating or partitioning the image into meaningful entities by defining boundaries between features and objects of an image based on intensity or texture criteria [72-74]. In medical image segmentation, prior knowledge of anatomy is used to identify the structures being considered [5, 75]. In the process of generating 3D models, image segmentation is a crucial step in determining the accuracy of the segmented region. The accuracy required of the segmented region varies depending on the purpose. For example, designing an anatomically pre-contoured fracture fixation plate does not require the same level of accuracy as is required for the finite element analysis of a bone model for stress analysis of an intramedullary nail-bone construct. In image segmentation, automatic processing is sometimes desirable as this can minimise operator involvement and reduce manual processing time. This is not always attainable due to the limitations imposed by the image acquisition and the complexity of the anatomical structures [73]. The articular regions of the long bones are often covered by a mixture of different types of tissues including cartilages, joint capsules, synovial fluid, ligaments, fat tissue, tendons and muscle. These different tissue types have different imaging properties and some of them are nearly impossible to be differentiated from the bone. For instance, the articular cartilage is not visible in CT images, while in MRI it is visible but often cannot be differentiated from the bone [15, 17]. Thus, depending on the segmentation method, considerable manual processing time is required to segment the articular regions of a long bone. Therefore, some of the segmentation methods are prone to operator introduced errors. Due to the difficulties of segmentation conferred by the partial volume effect and the complex anatomical structures, a large number of segmentation techniques that can 24

Chapter 3: Image processing and surface reconstruction

be used for segmentation of medical images of various anatomical regions have been reported [6, 76-82]. These vary from simple to complex methods that involve highly sophisticated mathematical algorithms as well as programming techniques [2, 4, 6, 76, 81, 83-85]. Among the various methods, manual segmentation and thresholding are relatively simple segmentation techniques commonly used for medical image segmentation compared to the region growing, artificial neural networks (ANN) and fuzzy logic based techniques where highly sophisticated programming and mathematics have been used.

3.3.1 Manual segmentation Manual segmentation is by far the simplest segmentation method available for medical image segmentation [83, 86]. The region of interest is manually delineated using a simple image editing or painting software program. Hence, there is no need for complex programming or software packages. The tracing of the ROI is usually carried out by a person with a good knowledge of both the anatomy of the desired region and image segmentation. This method is prone to inter- and intra-operator variability and the accuracy of the segmented region always depends on the knowledge and experience of the person who performs the segmentation [87]. The method is also more labour intensive and time consuming than the other segmentation methods available. Manual segmentation also has a poor repeatability compared to other segmentation methods and is not suitable for applications where high accuracy and repeatability is expected.

3.3.2 Intensity thresholding Intensity thresholding is a commonly used segmentation method for medical image segmentation that has been implemented in most of the commercially available image processing software packages [74, 88]. With this technique, the group of pixels (ROI) that has the intensity value above a set threshold level is assigned to one class while the rest (the background) is assigned to another class. Thus, a binary image that contains the ROI and the background is generated. In its basic form, this technique often relies on the user visually selecting a threshold level, thus making the method vulnerable to user dependent errors and less repeatable [89]. In addition, one threshold level does not accurately segment a complete long bone, as different regions of the bone have different intensity levels (Figure 3.1). 25

Chapter 3: Image processing and surface reconstruction

3.3.2.1 Selecting an appropriate threshold level The simplest method of selecting an appropriate threshold level for intensity thresholding is the visual selection of the threshold level. This is usually achieved by selecting a threshold level which reasonably selects the ROI without under- or overestimating it. Accuracy of the selected threshold level varies depending on several factors such as the window level setting of the display, and the knowledge and experience of the person. Therefore, this method is not appropriate in applications for which higher accuracy and repeatability are expected. Due to these drawbacks of visually selecting a threshold level and the unavailability of a standard method of selecting an appropriate level, various methods have been investigated [89]. Histogram based selection of threshold level [90-93] and clustering of grey levels of the boundary [94] are two of these methods. Most of the methods are not highly repeatable, and some involve complex programming that limits their use by a person with little knowledge of programming. Therefore, a repeatable and simple method of selecting a threshold level is required. 3.3.2.2 Multilevel thresholding Intensity thresholding is generally conducted using one threshold level to segment the complete bone or the region (global thresholding). However, global thresholding often fails to segment the complete bone accurately due to intensity inhomogeneity of the different regions of the bone [92, 95]. For example, the proximal, diaphyseal and distal regions of a long bone (femur or humerus) have different threshold levels, as illustrated in the graph below (Figure 3.1). Thus, the use of one threshold level to segment the complete bone will necessarily under- or overestimate the regions with different threshold levels.

26

Chapter 3: Image processing and surface reconstruction

Figure 3.1: Average intensity values of the outer bone contours as detected by the Canny filter for each axial CT image (350 slices) of the bone (The intensity values at the arrow locations were used to calculate the average thresholding values used for segmentation of each anatomical region) (HU = Hounsfield Units) The graph was obtained by plotting the threshold values calculated for each image slice against the slice number. The Canny edge detector based threshold selection method developed as a part of this research project was used to calculate the threshold value for each image slice. The graph shows that the threshold level for the diaphysis is fairly constant throughout its length, but the image slices of the proximal and distal regions have relatively low non steady threshold levels. For this reason, using a single threshold level to segment the complete bone will lead to inevitable inaccuracies of the segmented bone model. Thresholding the bone using more than one threshold level for regions with different threshold levels will segment the complete bone accurately and this has been successfully tested on small bones [95]. However, studies using multiple threshold levels for the segmentation of the complete long bones have not yet been reported in the literature and this will be investigated in the present study.

27

Chapter 3: Image processing and surface reconstruction

3.3.3 Edge detection Edge detection algorithms identify the rapid change of intensity level in a small neighbourhood of pixels in the image [96]. This is a popular segmentation method which has been used in cardiac and other medical image segmentation [76, 97]. As the algorithm considers local change of the intensity of a small region, this would be ideal for segmentation of an object with different intensity levels in different regions such as a long bone (e.g. femur). Positioning of the edge relative to the actual boundary of the object basically depends on the sensitivity of the edge detection algorithm used. Some of the algorithms allow users to change the sensitivity of the algorithm by choosing a threshold level. Edge detection has a higher repeatability compared to other segmentation methods as human intervention can be kept to a minimum level. However, this method is susceptible to artefacts and, more often, intensity changes due to noise are also detected as edges. The edge detection algorithms are necessarily complex programs; however, most of the algorithms are built into many image processing software packages (e.g. Matlab and IDL) and can be used easily. Among the number of edge detection algorithms available (such as Roberts and Sobel), Canny is an accurate, reliable and faster edge detector for image segmentation [98-101]. Therefore, the Canny edge detector was selected to investigate segmentation of long bones in the present study.

3.3.4 Region growing Region growing is a method of segmenting image regions or features that are connected, using pixel neighbourhood operations [102]. Starting from a user defined seed point, the region grows around it, extracting all the pixels connected to the seed point until the set criteria are met [74]. Intensity thresholding is often used in combination with region growing to segment the image features that are connected. It has also been used in skeletal system image segmentation [103]. The accuracy of the segmented region depends on the set criteria and this method often fails when used to segment complex structures such as long bones.

28

Chapter 3: Image processing and surface reconstruction

3.3.5 Sophisticated segmentation methods Most of the segmentation methods discussed above are basically implemented in 2D, in which each of the image slices is processed individually. Thus, a considerable amount of labour and time is required to segment a complete 3D volume. Therefore, fast and automatic segmentation methods have been investigated over the past few years. As a result, there are a number of robust image segmentation methods available for medical image segmentation. These methods carry out the segmentation process automatically, minimising the human intervention. In some of the techniques, simple methods such as intensity thresholding or edge detection have been used with modifications to automate the process, while other methods involve techniques such as artificial neural networks and fuzzy logic [6, 7, 81, 104-106]. These segmentation techniques utilise advanced programming techniques and mathematical algorithms, making them unavailable to the general research community with little knowledge of complex mathematics and advanced programming techniques. In addition, most of these techniques have been tested on smaller bones or part of a long bone which has relatively simple geometry compared to a human long bone [4, 107, 108]. Thus, these methods have the potential to fail or produce inaccurate results when used to segment a long bone with complex geometry, where image segmentation is particularly difficult due to restrictions imposed by image acquisition and anatomical structure variations. Therefore, further investigations using long bones are necessary before applying these methods on segmentation of long bones.

3.4 Surface generation The generation of triangular meshed 3D surfaces from the segmented boundary voxels is as important as the image segmentation, as the sub-voxel level accuracy of the 3D surfaces is mainly determined by this process [109]. This process also determines the number of triangles, their consistency, and their accuracy on the surface. The surface generation is usually carried out using one of the algorithms available [110, 111]. The marching cube algorithm (or its derivatives) is one of the popular algorithms that have been used in most of the commercially available software packages [62, 112]. In these packages, the surface generation is usually

29

Chapter 3: Image processing and surface reconstruction

carried out automatically, with the user being limited to setting the level of smoothing applied.

3.5 Registration (aligning) and comparison of surfaces Aligning of 3D surfaces is often used in research that involves 3D model manipulation. This aligning of the surfaces is first performed manually and then a surface matching algorithm is used to fine-align the surfaces. The iterative closest point (ICP) algorithm is a commonly used robust method for registration of 3D objects [1]. This algorithm has been used successfully in the literature with a high accuracy [113]. Lee et al. [1] conducted a registration test using the ICP algorithm in which a part of the bone model which had separated from the original model was matched perfectly to its original full model. The algorithm has also been implemented in most of the 3D modelling software packages and, therefore, is easily accessible. After the registration of the 3D surfaces, the comparison is usually carried out by calculating the deviation of the surface of interest from the reference surface. This is conducted using a point to point comparison method where the normal distance from a point of the surface of interest to a corresponding point of the reference surface is calculated.

3.6 A reference standard for validating 3D models of bones Validation of 3D models plays an important role in the studies that involve 3D models of bones, especially when live subjects have been used, where the physical bone is not readily accessible. A number of methods have been established over time; however, none of these is accepted as a standard method for validation of 3D models of bones. Amongst the methods used, models manually segmented by anatomy experts have been used to validate the 3D models [4]; however, the accuracy of this method is highly dependent on the experience and knowledge of the person who conducts the segmentation. 3D laser scanning of the bone‟s surface has been used in several studies [3, 69, 83, 114]. In this method, an outer coating has been applied on the bone‟s surface; however, this coating might introduce errors to the scanned surface unless applied evenly. Mechanical digitisers or digitising arms are other options for 30

Chapter 3: Image processing and surface reconstruction

digitising the bone surface. There are studies that report on the use of mechanical digitising arms; however, these devices are not capable of generating an evenly distributed mesh unless they move automatically [115]. One of the important limitations of laser and mechanical digitisers is that none of these methods can be used to validate the internal medullary canal of a long bone. As the medullary canal is 1-2 cm in diameter, the scanner head of a digital scanning arm or the laser scanner cannot reach the inside of this canal. Even though scanning of the cut-opened bone canal is possible, the bone loss from the bone saw (0.5 -1.0 mm) is inevitable and this can lead to errors in the final 3D model. In the present study, an attempt was also made to model the medullary canal using dental acrylic (PMMA); however, this was not possible as the material shrinks when it solidifies and also generates air bubbles, thus causing some regions of the PMMA mould to lose contact with the bone. Goyal et al. used MicroScribe digitiser–a mechanical arm with a stylus–to capture 3D points from tibial surface fitted with a plate [46]. It has an accuracy of up to 0.23 mm and sampling rate of 1000Hz. The study used the scanning arm only to record the position of the plate and tibiae and did not generate the complete 3D model. Gelaude et al. used a laser strip scanner which measures the distance to an object from the scanner head [3]. In combination with a coordinate measuring system, this was used to generate a reference standard for the soft tissue free human femora, obtaining an accuracy of 0.70 ± 0.55 mm when compared with CT derived models of the same samples. DeVries et al. also used a laser scanner (Roland LPX-250) to validate phalanx 3D models. The scanner was used with a resolution of 0.2 mm and there was an average 0.2 mm deviation from the manually segmented CT based 3D models. Considering the advantages and disadvantages of the methods described, the following two methods were used to validate the 3D models of long bones reconstructed from CT and MRI data of ovine femora. A contact mechanical scanner, which automatically moves along the object being scanned, is a good option for accurate digitisation of a denuded bone‟s outer surface. The scanner moves automatically along a pre-defined mesh, essentially generating an evenly distributed mesh that cannot be achieved with mechanical digitising arms. A MicroCT scanner is 31

Chapter 3: Image processing and surface reconstruction

capable of scanning an object with a very high resolution (e.g. 30 µm isotropic voxel size). This also has the ability to digitise the inner medullary canal, as well as complex geometric shapes that cannot be digitised with methods such as laser or mechanical digitising arms. Therefore, this is an ideal method for the validation of the medullary canal of long bones.

3.7 Aims of the study This study specifically aimed to: Investigate the accuracy of multilevel intensity thresholding as a method of segmenting CT data of long bones in combination with a new threshold level selection method Investigate the accuracy of Canny edge detection for segmentation of CT data of long bones Compare the accuracy of multilevel intensity thresholding and Canny edge detection to single level thresholding

3.8 Methods 3.8.1 Samples Five intact cadaveric sheep hind limbs, amputated from the pelvis, were obtained from four Merino-Cross sheep. However, the statistical analysis of the sample size for 80% power shows that 28 samples are needed to detect a difference of 0.06 mm with SD = 0.04 and 0.015. Due to the long processing time of the samples, it is not practicable to use these sample sizes and therefore, a sample size of 5 has been used. Using the sample size of 5 the difference that can be detected is 0.108 mm with the same standard deviations.

3.8.2 Image segmentation Three image segmentation methods were investigated in this research project for segmentation of CT data: single-level intensity thresholding; multilevel intensity thresholding and Canny edge detection. Single level thresholding was performed for the purpose of comparing this method with the other two methods. Multilevel thresholding was used to overcome the problem of over- or underestimating the regions with different threshold levels when a single threshold level is 32

Chapter 3: Image processing and surface reconstruction

used. A method of selecting an appropriate threshold level was also used with the multilevel thresholding. This threshold selection method was used to reduce the user dependent errors of visually selecting a threshold level. The threshold selection method was based on the calculation of average intensity of the edge that is detected by the Canny edge detection filter. In multilevel thresholding, the Canny edge detector was used to determine the threshold level utilising Canny edge detector‟s higher repeatability as a superior method to visual selection of the threshold level. Thus, the multilevel thresholding and Canny edge detector methods were expected to be similarly accurate for segmentation of CT image data. The Canny edge detection filter was used to delineate the outer and inner cortex from the bone as the third segmentation method. Canny edge detection was performed in 2D axial images, and then the outer and inner edges of the bone cortex were delineated using a customised Matlab script. These edges were later combined to reconstruct the 3D models of the outer and inner cortex of the femur. A detailed section of the segmentation methods used in this part of the research project is available in the paper presented at the end of this chapter.

3.8.3 Reference model for validation of the outer 3D models Validation of the outer 3D models was carried out using a contact mechanical scanner (MDX-20 Roland) to digitise the outer surface of the bone. The complete process involved the prior removal of the soft tissues from the bone and then scanning of the outer surface in several steps, generating a number of surfaces. Finally, the reference 3D model was reconstructed by merging the scanned surfaces. 3.8.3.1 Removal of the soft tissues from long bones Various methods have been reportedly used for removing soft tissues from bones, such as boiling or use of chemicals to dissolve the tissues [3]. These methods have the risk of changing the outer geometry of the bone and therefore were not used in this study. The removal of soft tissues before the scanning with the contact mechanical scanner was achieved by dissecting the limb with a scalpel. After the bone has been harvested, the scalpel blade was used to carefully remove the soft tissue attached to the bone, without damaging the bone‟s outer geometry (Figure 3.2). 33

Chapter 3: Image processing and surface reconstruction

Figure 3.2: The process of removing soft tissues from the sheep femur before scanning with the contact mechanical scanner: a - gross dissection with the scalpel, b - removing soft tissues attached to the bone, and c - soft tissue free bone 3.8.3.2 Scanning of the bone’s outer surface using the contact scanner A mechanical 3D contact scanner (Roland DG Corporation, MDX 20, Japan) (Figure 3.3) was used to digitise the surface of the denuded bone. The MDX 20 scanner scans an object in the horizontal plane (x, y Plane), moving its head on “x” direction while moving the stage in “y” direction. A needle connected to the head containing an active piezo sensor moves vertically (z direction) perpendicular to the x y plane until it touches the surface of the object and records the x, y and z coordinates of the position of the needle. Then, the head moves towards x direction at a set distance and records the position of the needle. Finally, the scanner collects a point cloud with an x, y and z coordinate for each point. The manufacturer‟s specifications of the scanner are given below (Table 3.1). The active piezo sensor, to which the needle is connected, is highly sensitive and ensures that the needle stops before it damages the surface of the object being scanned.

34

Chapter 3: Image processing and surface reconstruction

Figure 3.3: Scanning of the bone's outer surface of the diaphyseal region using the MDX 20 contact scanner (The bone is positioned on the stage using glue tags) Table 3.1 Specifications of the MDX 20 contact 3D scanner3 Property Sensor Scanning method Scanning pitch Scanning speed Exportable file formats XY table size

Value Roland Active Piezo Sensor (R.A.P.S.) Probe length 60 mm (2-5/16 in.), tip bulb diameter 0.08 mm (0.00315 in.) Contacting, mesh-point height-sensing X/Y-axis directions -0.05 to 5.00 mm (0.002 to 0.20 in.) (Settable in steps of 0.05 mm (0.002 in.)) Z-axis direction - 0.025 mm (0.000984 in.) 4-15 mm/sec. (1/8-9/16 in./sec.) DXF, VRML, STL, 3DMF, IGES, Greyscale, Point Group and BMP 220 (X) x 160 (Y) mm ( 8-5/8 x 6-1/4 in.)

Dr.PICZA software package (Intellecta Technology Pty Ltd, Adelaide, Australia) installed on a personal computer was used for the operation of the scanner and for the acquisition of the x, y and z coordinates from the scanner. In the present study, the bone outer surface was digitised with a resolution of 0.3 mm × 0.3 mm in the scanning plane (x, y plane) and a step size of 0.025 mm in the vertical direction (z direction). The scanning of the surfaces of the bones was performed in two stages, scanning of the diaphysis and the articular regions.

3

The information was drawn from the manufacturer‟s website.

35

Chapter 3: Image processing and surface reconstruction

To digitise the diaphysis of the bone, the soft tissue free bone was positioned horizontally on the stage of the scanner (Figure 3.3). The bone was firmly fixed on the stage using two sets of glue tags (Figure 3.3). Mechanical clamps were not used so as to prevent damage to the bone‟s surface. Scanning of the shaft was carried out in five steps, rotating the bone around its long axis by approximately 70° to generate a total of five surfaces. Five surfaces were used in order to keep a good overlap between two consecutive surfaces for their alignment using an ICP algorithm based method. The bone was then cut into three parts in order to scan the articular surfaces as these regions cannot be reached while the bone is intact (Figure 3.4). The bone was divided such that the distal and proximal parts were not longer than 50 mm, as the maximum height of an object that can be scanned by the scanner is 55 mm. Then, the proximal or distal part was positioned vertically on the stage to scan the articular surfaces (Figure 3.5). The number of scans required for completely digitising the bone‟s articular surfaces was determined by the complexity of the geometry of these surfaces. In general, five to ten scans were carried out in each of the articular surfaces. Before the scanning of articular surfaces started, the 3D model of the shaft was reconstructed (Figure 3.6) and used as a guide to locate the areas to be scanned. The scanned surfaces were exported as STL files for further reconstructing the 3D model in Rapidform 2006.

Figure 3.4: Bone is cut in three parts in order to scan the articular surfaces which cannot be reached by the scanner on the intact bone

36

Chapter 3: Image processing and surface reconstruction

Figure 3.5: Positioning of the proximal articular segment of the femur in order to scan the articular surface

Figure 3.6: The reconstructed model before the scanning of articular surfaces (This model was used as a guide to scan the articular regions) 3.8.3.3 Reconstruction of the 3D model from scanned surfaces The 3D reference model was reconstructed from the scanned surfaces using the reverse engineering software package Rapidform 2006. This was carried out in three steps: 1. Removal of unusable data from the scanned surfaces 2. Aligning of the consecutive surfaces and 3. Merging of the aligned surfaces to generate the final 3D model.

37

Chapter 3: Image processing and surface reconstruction

As the first step, unusable triangles were removed from the scanned surfaces. The scanner could acquire geometric data only on the horizontal plane (Figure 3.7 & Figure 3.8). Thus, the geometric data in the horizontal plane is displayed with equilateral triangles, while unusable data is usually represented by triangles which have long faces. As a result, the surfaces contained triangles with variable length. The triangles which are longer than 1 mm (roughly) were determined to be the unusable data. Using functions built into Rapidform 2006, these unusable triangles were removed permanently from the surfaces. In addition, the triangles which made up the parts of the stage and the blue tags used to hold the sample to the stage were also removed from the surfaces.

Figure 3.7: Scanned surface with unusable data

Figure 3.8: The surface after removing the unusable data In the second step, two cleaned consecutive surfaces were aligned using the methods described in Section 3.8.4.1(Figure 3.9). Once the alignment of the first two surfaces

38

Chapter 3: Image processing and surface reconstruction

was performed, the second surface was kept locked and the third surface was aligned to the second.

Figure 3.9: Two adjacent surfaces are fine registered Finally, all aligned surfaces were merged using the „Merge Surfaces‟ function built into Rapidform 2006 (Figure 3.10). Once the merging of the surfaces was completed, re-meshing of the triangular surface was carried out to obtain equilateral triangles and an evenly distributed mesh.

Figure 3.10: The final 3D model reconstructed by merging the surfaces

3.8.4 Reference model for validation of the medullary canal As the contact or laser scanner is unable to reach the medullary canal of ovine femora, a microCT scanner was used to generate a reference standard. The scanning was conducted only for the diaphyseal region due to the limitation of the sample length that the scanner could accommodate. The medullary canal of the soft tissue free bone diaphysis (Figure 3.4) was cleaned to completely remove the bone marrow using a detergent solution and a brush. The bone marrow was removed in order to have equal interfaces (bone-water) in the outer and inner bone cortex. The bone was immersed in pure water and scanned with the microCT (microCT 40, Scanco medical, Switzerland) scanner using the scanning protocol shown in Table 3.2. Segmentation of the image data was conducted using 39

Chapter 3: Image processing and surface reconstruction

the Canny edge detection method described in Section 3.8.1, generating 3D models of the outer and inner surfaces. Before the segmentation with the Canny filter, a 20 × 20 median filter was applied to reduce the salt and pepper noise contained in the images (Figure 3.11). Table 3.2 Scanner parameters used for microCT scanning Parameter Resolution Slice spacing kVp

Value 0.03 mm × 0.03 mm 0.03 mm 140

Figure 3.11: a - The original microCT image (a cross section from the diaphysis); and b - the image after applying a 20 × 20 median filter Both microCT based outer and inner 3D models were subjected to 90% decimation to reduce the number of triangles that were contained within the 3D models to 900 000. As a result of 0.03 mm voxel size used for microCT scanning, the final 3D models contained about 9 000 000 triangles; this made the models difficult to handle in the software systems used for the study. The number of triangles contained in the microCT based models after the decimation was still higher than the number of triangles contained within the contact scanner generated reference models (~350 000); this indicated that the microCT based model was accurate enough to use as a reference standard. The outer 3D models generated from microCT images were validated with the contact scanner generated 3D models, resulting in a nearly uniform error of 0.12 mm, where the microCT model underestimated the reference model. As there was no 40

Chapter 3: Image processing and surface reconstruction

significant deviation of the outer 3D models from the reference standard, it was assumed that the microCT generated inner models could be used to validate the inner models generated from the CT and MRI data.

3.8.5 Basic 3D modelling techniques using Rapidform 2006 Throughout this study, the Rapidform 2006 (INUS Technology inc. Korea) reverse engineering software package was used for the reconstruction and manipulation of 3D models. Registration of a model of interest to the reference standard was conducted using a built in function that is based on the ICP algorithm. The comparison of the model of interest to the reference model was conducted using a point to point comparison method available in Rapidform 2006 software system. 3.8.5.1 Registration of 3D surfaces using Rapidform 2006 Aligning of surfaces was basically used on three occasions in this study: first, in the aligning of the contact scanner generated surfaces of the denuded bone in order to generate the reference model; second, in the aligning of 3D models prior to the quantification of the geometric deviation between a model of interest and the reference model; and, third, in the correction of the lateral shift artefact of 3D models of long bones based on MRI. The registration process of the surfaces or models was carried out in two steps: gross alignment and fine alignment of the surfaces. The gross alignment was accomplished using the „Shell Trackball‟ function built into Rapidform 2006. The reference model was locked in 3D space to prevent accidentally moving the model. The model of interest was then connected with the trackball and moved until the model was roughly in alignment with the reference model (Figure 3.12). The trackball tool allows moving a 3D model in x, y and z directions and rotating around those three axes using the mouse. The fine alignment of the models roughly aligned with the trackball was carried out using the built in function „Fine Registration‟ which is based on the ICP algorithm.

41

Chapter 3: Image processing and surface reconstruction

Figure 3.12: The initial aligning of the CT based 3D model to the reference model using Trackball prior to the application of fine registration function

Figure 3.13: A CT based model (red) is aligned to the reference model (blue) in Rapidform 2006 using the fine registration function 42

Chapter 3: Image processing and surface reconstruction

3.8.5.2 Comparison of the aligned 3D models A method of calculating average displacement between two surfaces was used for the comparison of a model of interest to the reference model [116]. This method calculates the average of the deviations of the points in the model of interest to the corresponding points in the reference model. The method is built into the Rapidform 2006 software package and was used on the surfaces that had been aligned using the method described in the previous section. A graphical representation of the distribution of the point to point deviations was also generated by the software package (Figure 3.14). The 3D models were compared as complete models; however, in some of the investigations (Chapters 3 and 5), the different anatomical regions of the models were also compared in order to quantify the errors associated with each anatomical region (Figure 3.15).

Figure 3.14: Comparison of the aligned CT model to the reference model in Rapidform 2006

43

Chapter 3: Image processing and surface reconstruction

Figure 3.15: Five anatomical regions used for the comparison: 1 - femoral head, 2 proximal region, 3 - diaphysis, 4 - distal region, 5 - distal articular region 3.8.5.3 Dividing the 3D models of bones into different anatomical regions Where the comparison of different anatomical regions was required, the 3D models were divided into five anatomical regions (Figure 3.15) according to the guidelines given in „AO principles of fracture management‟ [117]. The bone was divided using two reference planes and two curves created in 3D space of Rapidform 2006 (Figure 3.16). The same reference planes and curves were used to divide all the models of one sample.

Figure 3.16: Reference planes and curves used for the splitting of the model into five anatomical regions

3.9 Results Comparison of the complete outer bone models based on three segmentation methods to the reference model generated average deviations of 0.24 mm, 0.18 mm and 0.20 mm for single threshold, multi-threshold and edge detector methods respectively. Comparison of inner medullary canal models generated average deviations of 0.43 mm for the single threshold method, 0.17 mm for the multi-threshold method and

44

Chapter 3: Image processing and surface reconstruction

0.27 mm for the edge detector method. Detailed results are available in the paper presented at the end of this chapter.

3.10 Summary, discussion and conclusion Increased utilisation of virtual 3D models of long bones for various practices in the clinic and in research has made the 3D reconstruction of long bones a research interest. The process of reconstructing a 3D model involves several steps and each step has factors that determine the geometric accuracy of these models. Among these steps, the data acquisition, segmentation and surface reconstruction are equally important; however, image segmentation is the mostly user intervened process and has been discussed widely in the literature. While a large number of methods are available for segmentation of bone data from CT or MRI data, intensity thresholding is the most commonly used method due to its ease of use. The unavailability of a method to select the appropriate threshold level means that this method relies mainly on visual selection of a threshold level. In addition, a single threshold level does not accurately select the ROI from all the anatomical regions of a long bone, as different regions require different threshold levels. The Canny edge detector is another segmentation method that can be easily implemented as it is already incorporated in many of the image processing software packages (e.g. Matlab). However, in the relevant literature, there is no reported use of the Canny edge detector for segmentation of long bones. In the present study, intensity thresholding and the Canny edge detector were investigated for their accuracy and repeatability in segmenting the CT data of long bone from ovine hind limbs. These two methods were selected as they do not involve complex programming and can be administered by researchers with a limited knowledge of programming and mathematics. A threshold selection method based on the Canny edge detector was introduced for intensity thresholding to minimise the user dependent errors of selecting a threshold level. In addition, a multilevel thresholding approach was used instead of a single threshold level for segmenting the complete long bone. Intensity thresholding with a visually selected single threshold was also carried out for comparison purposes.

45

Chapter 3: Image processing and surface reconstruction

The results indicate that the multilevel intensity thresholding approach with the threshold selecting method can produce 3D models with a relatively higher accuracy (average deviation = 0.18 mm), in comparison to edge detection (average deviation = 0.20 mm) and the single threshold method (average deviation = 0.24 mm). However, the overall accuracy obtained from all three methods was within acceptable range (0.18 – 0.24 mm) for reconstruction of accurate 3D models, depending on the accuracy required by the specific application. When different anatomical regions are considered, the multi-threshold method was able to generate accurate models for most of the regions, while single threshold generated the least accurate models for most of the regions. Compared to the single threshold method, the other two segmentation methods had a relatively higher repeatability. The study utilised 3D surfaces derived by mechanically digitising the denuded bone surfaces for an accurate validation of CT based models. This method is also used for the validation of MRI based 3D models in next part of the research. A limitation of this method was that no measures were considered for preventing the dehydration of the bones during the digitisation. Practically this was difficult to achieve as the bone‟s surfaces could not be covered during the scanning. There is no evidence to suggest that dehydration has an effect on the cortical bone geometry; however, this shrinks the cartilages which might be a reason for higher error occurred in this region. The number of samples used was also limited to five due to longer processing times even though the calculated sample size was 28 to detect the obtained difference. With the sample size of five the detectable difference is 0.108 mm. The accuracy required for designing orthopaedic implants are in the order of few millimetres and thus, a difference of 0.108 mm would not affect the accuracy of the reconstructed models. This study demonstrated that by using relatively simple segmentation methods, 3D models with sub-voxel accuracy can be generated. This allows the general research community to use relatively simple methods without having to involve complex programming and mathematics. The segmentation methods investigated in this part of the research project will be used to segment the CT and MRI bone data throughout the project.

46

Chapter 3: Image processing and surface reconstruction

The next chapter demonstrates an application of 3D models generated from CT data where a validation of two intramedullary nail designs was conducted in a 3D environment using 3D models of nails and the intramedullary canal of the tibia. The study utilised 3D models based on CT scans of cadaver bones, however, if MRI is used for scanning, this method can be used to assess the fit of intramedullary nails to patient‟s bones.

47

Chapter 3: Image processing and surface reconstruction

3.11 Paper 1: Effect of CT image segmentation methods on the accuracy of long bone 3D reconstructions (published)

48

Chapter 3: Image processing and surface reconstruction

226-233

Contents lists availabl e at ScienceDirect

Medical Engineering & Physics journal homepage: www.elsevier.com/locate / medengphy

ELSEVJER

Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions Kanchana Rathnayaka a, Tony Sahama a, Michael A. Schuetz a.b, Beat Schmutza.• • Institute of Health and Biomedical Innovation. Queensland University of Technology, Brisbane, Australia b Department of Orthopaedics. Princess Alexan dra Hospital. Brisbane. Australia

ARTICLE

INF O

Article history:

Due to copyright restrictions, this article is not ABSTRACT available here. Please consult the hardcopy thesis available from QUT Library or view the published An accurate and accessible image segmentation method is in high demand for generating 3D bone models fromat: CT scan data, as such models are required in many areas of medical research. Even though numerous version online

Received 6 May 2010 Received in revised form 20 August 2010 Accepted 4 October 201 0 Keywords:

Computed tomography Image segmentation Canny edge detection Thresholding Bone models MicroCT Femur Mechanical digitiser

sophisticated segmentation methods have been published over the years, most of th em are not readily available to the ge neral research community. Therefore. this study aimed to quantify the accuracy of three popular image segmentation methods, two imp lementations of intensity thresholding and Canny edge detection, for generating 3D models of long bones. In order to reduce user dependent errors a ssociated with visually se lecting a threshold value, we present a new approach of selecting an appropriate threshold value based on the Canny filter. A mechanical contact scanner in conjunction w it h a microCT scanner was utilised to generate the reference models for validating the 3D bone models generated from CT data of five intact ovine hind limbs. When the overall accuracy of the bone m odel is considered, t he three investigated segmentation me thods generated comparable results w ith mean errors in the range of 0.18- 0.24mm. However, for t he bone diaphysis, Canny edge detection and Canny filter based thresholding generated 3D models w it h a significantly higher accuracy compared to those generated through visually selected thresholds. This study demonstrates that 3D models with sub-voxel accuracy can be generated utilising relatively simple segmentation methods that are available to the general research community. © 2010 IPEM. Published by Elsevier Ltd. All rights reserved.

http://dx.doi.org/10.1016/j.medengphy.2010.10.002

1. Introduction Accurate three-dimensional (3D) models of long bones are required in appl ications, such as implant design [ 1-5). finite element analysis (FEA) [6-11) and computer-aided surgical planning [ 12-15]. Computed tomography (CT) is curren tly the gold standard for the acquisition of data from wh ich the 3D models of long bones are generated. Two main steps are involved in gene rating a 3D model from aCT data set: image segmentation; and 3D reconstruction of the segmented bone contours. In commercial image data processing and 3D reconstruction packages the latter is performed automatically with t he user being limited to choose the level of surface smoothing to be applied. While surface smoothing can influence the accuracy of the reconstructed bone model (16,17]. for the purpose of this study smoothing was treated as a fixed entity (default setting of the commercial software package). Therefore, it is the former that was investigated, as an accurate and reproducible image segme ntation method is a necessity for gener-

* Corresponding author at: Institute of Health and Biomedical Innovation, 60 Musk Avenue, Kelvin Grove. QLD 4059, Australia. Tel.: +61 7 3138 6238: fa x: +61 7 3138 6030. £-mail addresses: [email protected], [email protected] (B. Schmutz).

ating 3D models that are accurate geometric representations of the actual bones. Segmentation techniques are used to separate the region of interest (ROI ) from the remainder of the image. The segmentation is critical as it is the major step demarcating between ROl and t he background and thus. has a major effect on the geometric accuracy of the 3D model. Therefore, studies have been carried out to develop segmentation techniques that can produce 3D models with a high geometric accuracy. As a result, many image segmentation methods are available ranging from manual segmentation to semi and fully a utomated techniques (17-24]. Manual segmentation/traci ng of the ROI by humans has long been practiced (17] and is so far the simplest method available for medical image segmentation. The major disadvantages of the manual segmentation are intra- and inter-personal variability which makes it a less repeatable method. This method is also more labour intensive and time consumi ng than the other segmentation techniques available. Intensity thresholding is a popular segmentation method, which is implemented in commercial medical Image 3D reconstruction packages. In its basic form, this technique relies on visual selection of the t hreshold level by the user which has an effect on the accuracy and the repeatability of this method. In the absence of a standard method of selecting an a ppropriate threshold level, various

1350-4533/$ - see front matter © 2010 IPEM. Published by Elsevier Ltd. All rights reserved. doi:10.1 016/j.medengphy.201 0.1 0.002

49

Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation

Chapter 4 Application of 3D modelling techniques for orthopaedic implant design and validation 4.1 Introduction Three dimensional models (3D) with accurate geometric representation of long bones are increasingly being used for various aspects of clinical practice and research. They provide a useful platform for the design and validation of implants, avoiding the necessity to use cadaver bones. They also provide researchers with an opportunity to design and validate implants for younger age groups who are more prone to injuries and for whom there are only a few cadavers available. Designing implants that fit the anatomy of young age and ethnic groups is also of particular important as age and ethnicity are two of the factors that determine the geometric and mechanical properties of long bones. Even though 3D models have been used for implant design, their use for validation of the anatomical fit of the implants has seldom been reported. Therefore, to address this need, this study investigates an in-silico validation process of two intramedullary nail designs using triangular meshed 3D surfaces generated from CT data of cadaver bones. The reconstruction process of accurate 3D models from CT data is discussed in detail in Chapter 3 of this thesis. This chapter now focuses on the application of these models in the validation process of already designed implants. Section 4.2 discusses the relevant literature and Section 4.4 briefly introduces the methodology used. A detailed methods section is available in the journal paper presented at the end of this chapter.

57

Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation

4.2 3D models for implant design and validation The conventional use of cadaver bones for implant design and validation has a number of challenges that researchers have to face. Most of the available cadaver bones are basically obtained from older (>60 years old) donors. Thus, these cadaver bones do not represent the young patient population who make up about half of the patients who require implants. As most of the implant validation studies are carried out in regions where fewer cadavers of Asian origin are available, the use of such cadavers for implant design and validation is limited. Anecdotal clinical evidence also suggests that the currently used trauma fixation plates do not optimally fit the bones of patients from the Asia-Pacific region, as they have been designed mainly for the Caucasian population.[45] Therefore, the implant design and validation process needs to be extended to both the young and Asian-pacific population. Accurate 3D models of small or long bones provide a better platform for design and validation of implants for different age and ethnic groups. Using MR imaging, 3D models of long bones can be reconstructed from almost all age groups, as MRI is a potential alternative to CT for generating 3D models of bones (See Chapter 3). The limitation of not having enough Asian-Pacific cadavers can also be overcome by using 3D models generated from such populations using MRI. This use of MRI also enables researchers to repeatedly use the same specimen for validation studies in a simulated environment without having to damage the already available, valuable cadaver specimens. There are only a few studies that have been conducted to quantify the anatomical fit of an implant. The first reported is the study conducted by Goyal et al. [46] using 101 tibiae and medial and lateral proximal periarticular plates. In this study, the quantification of the fit was conducted by digitising the position of the plate and the bone, using a mechanical digitising arm (Microscribe). Haraguchi et al. [118] used CT scans of 50 patients and a ORTHODOC workstation to compare the fit and fill between anatomic stem and straight tapered stem. This was performed using 3D surfaces extracted using the software system and placing the implants virtually in the 3D surface models. A study quantifying the plate fit using 3D models has been reported by Schmutz et al. [45]. Twenty one 3D models of tibiae from a database at AO Development Institute, Davos, Switzerland and a 3D model of the distal 58

Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation

periarticular tibia plate were used. Using 3D modelling techniques, the distance between plate and the bone and the angle between plate and the bone were measured to assess the fit of the plate to the bone. Even though these few studies on validation of plates have been reported, no studies to validate intramedullary nails using 3D modelling techniques have been conducted to date. The insertion force and insertion distance of the nail is often used as an indicator for anatomical fitting of a nail; however, fit of an intra-medullary nail in the final position cannot be quantified using these methods. The validation using cadavers is also limited by the small number of available cadavers, and those that are available might not be representative of the target population‟s age and ethnicity.

4.3 Aims of the study This study aims to develop a non-invasive method to quantitatively assess the anatomical fitting between an intramedullary nail in the final position and the bone, using 3D models of long bones.

4.4 Methods The study used two designs of the expert tibial nail (ETN): ETN and ETN with bend (Synthes, Bettlach, Switzerland) and 20 CT based 3D cortex models of Japanese cadaver tibiae. 3D models of the ETN and ETN with bend were virtually positioned in the 3D model of tibiae using the Rapidform 2006 software system to meet the set criteria to obtain the optimal fit. The maximum distance and the area of the part of the nail protrusion were measured using 3D modelling techniques. A detailed methodology used for the study is available in the paper presented at the end of this chapter.

4.5 Results The total area of the nail protruding from the medullary canal was 540 mm2 for the ETN with bend, and 1044 mm2 for the ETN. The maximum distance of the nail protruding from the medullary canal was 1.2 mm for the ETN with bend, and 2.7 mm for the ETN.

59

Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation

4.6 Summary, discussion and conclusion The limited accessibility to cadaver bones of younger age groups and different ethnicity makes the design and validation of the implants specific for them a difficult process. The use of cadaver bones allows assessment of nail insertion force; however, available cadavers are limited in number and do not represent the target population. Validation using plain x-ray is also limited to 2D. The use of 3D models provides the opportunity to access bone geometric data from younger age groups of different ethnicity, using non-invasive MRI scanning. This also allows for repeated use of the same sample for implant validation, and provides an accurate method to quantify the anatomical fit of implants. This study quantified the anatomical fit of two nail designs (ETN and ETN with bend) to the 3D models of tibiae reconstructed from CT data of Japanese cadavers. Based on the results, the total area and the maximum distance of the nail protruding from the medullary canal were smaller for the ETN with bend compared to the ETN. Both protruding area and the distance showed statistically significant differences between ETN with bend and ETN. Therefore, compared to the original ETN, the modified nail design (ETN with bend) had a better fit. This will provide a better alignment of the fractured bone segments, resulting in a better fracture healing outcome. The method presented in the study using 3D models of the nails and tibiae was noninvasive. This is also radiation hazard free when MRI is used to scan the bones. Thus, this method has the potential of validating the nails or plate designs for healthy human volunteers who represent the target patient populations of young age and different ethnicity.

60

Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation

4.7 Paper 2: Quantitative fit assessment of tibial nail designs using 3D computer modelling (published)

61

Chapter 4: Application of 3D modelling techniques for orthopaedic implant design and validation Injury, lnt. j . Care Injured 41 (2010) 216-2 19

Contents lists available at ScienceDirect

Injury ELSEVIER

journal h omepage: www.elsevier.com / lo catelin ju r y

Quantitative fit assessment of tibial nail designs using 30 computer modelling B. Schmutz a.*. K. Rathnayaka a. M.E. Wullschleger a.b,]. Meek c. M.A. Schuetz a,b • Institute of Health and Biomedical Innovation, Queensland University of Technology, 60 Musk Avenue, Kelvin Grove, Brisbane, QLD 4059, Australia b Trauma Services, Princess Alexandra Hospital, Brisbane, Australia c Synthes GmbH, Oberdorf, Switzerland

ART I CLE I NFO Article history: Accepted 5 October 2009 Keywords: 3D model Tibia Intramedullary nail Nail fit Fracture fixation

Due to copyright restrictions, this article is ABST R ACT not available here. Please consult the Intramedullary nailing isavailable the standard fixation displaced diaphyseal fractures of the tibia in hardcopy thesis from method QUT for Library adults. The bends in modern tibial nails allow for an easier insertion, enhance the 'bone-nail construct' orstability, view and thereduce published versionof the online at: axial malalignments main fragments. Anecdotal clinical evidence indicates that current nail designs do not fit optimally fo r patients of Asian origin. The aim of this study was to develop a method to quantitatively assess the anatomical fitting of two different nail designs for Asian http://dx.doi.org/10.1016/j.injury.2009.10.012 tibiae by utilising 3D computer modelling. We used 3 D models of two diffe rent tibial nail designs (ETN (Expert Tibia Nail) and ErN-ProximalBend, Synthes), and 20 er-based 3D cortex models ofJapanese cadaver tibiae. With the aid of computer graphical methods, the 3D nail models were positioned inside the medullary cavity of the intact 3D tibia models. The anatomical fitting between nail and bone was assessed by the extent of the nail protrusion from the medullary cavity into the cortical bone, in a real bone this might lead to axial malalignments of the main fragments. The fi tting was quantified in terms of the total surface area, and the maximum distance by which the nail was protruding into the cortex of the virtual bone model. In all 20 bone models, the total area of the nail protruding from the medullary cavity was smaller for the ETN-Proximai-Bend (average 540 mm 2 ) compared to the ETN (average 1044 mm 2 ) . Also. the maximum distance of the nail protruding from the medullary cavity was smaller for the ErN-ProximalBend (average 1.2 mm) compared to the ETN (average 2.7 mm). The di fferences were statistically significant (p < 0.05 ) for both the total surface area and the maximum distance measurements. By utilising computer graphical methods it was possible to conduct a quantitative fit assessment of different nail designs. The ETN-Proximai-Bend shows a statistical significantly better intramedullary fi t with less cortical protrusion than the original ETN. In addition to the application in implant design, the developed method could potentially be suitable for pre-operative planning enabling the surgeon to choose the most appropriate nail design for a particular patient. © 2009 Elsevier Ltd. All rights reserved.

Introduction Intramedullary nailing is the standard fi xation method for dis placed dia physeal fractures of the tibia in adults. 6 ·10 The bends in mode rn tibial nails a llow for a n easie r insertion, e nhance the 'bone-nail construct' stability, and reduce axial malalignments of the main fragme nts.3 - 5 ·9 Typically, the nails a re designed with the view to fit the 50th percentile of a Caucasian/W este rn population. Clinical tria ls of the nail designs a re then conducted in hospitals were the majority of the patients are of Caucasian orig in. Such was the case for the Expert Tibial Nail (ETN), one of the nail design s used in th is study. The results of a clinical study13 from one of the

• Corresponding author. Tel.: +61 7 3138 6238; fax: +61 7 3138 6030. £-mail address: [email protected] (B. Schmutz). 0020- 1383/$ - see front matter ~ 2009 Elsevier Ltd. All rights reserved. doi:10.101 6fj.injury.2009.1 0.012

62

hospitals involved in the multi-centre clinical tria l o f this nail confirmed the improvements 5 of the nail design as appropriate fo r their patient collective. exclusively of Caucasian origin (Striegel A, personal communication, July 19, 2009 ). Despite this. anecdotal clinical feedback is emerging, indicating that the curre nt nail des ign does not fit optimally in the proximal dorsal region for the tibial geometry of As ian patients. One important aspect of d esigning a new or improved implant s hape is validation, which is often conducted in the form of cadaver trials. In the case of precontoured plates, the anatomical fitting can be visually assessed or quantified by fitting plates to bones.2 However. for nails, one aspect of validation pertains to the ease w ith w hich the nail can be inserted into the bone, and the other to the anatomical fitting between the nail and bone geometry in the nail's final position. Neither of these can be achieved in form of a vis ual assessment. The force required for inserting the nail into the bone is

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

Chapter 5 Magnetic resonance imaging for 3D reconstruction of long bones 5.1 Introduction The widely accepted standard for generating 3D models of bones for implant design and related research is CT as this offers a better contrast at the bone–soft tissue interface that greatly facilitates the segmentation process. However, due to the involvement of a high dose of ionising radiation, CT cannot be used for scanning of healthy volunteers for research purposes. Therefore, an alternative radiation free imaging method such as MRI is necessary for the generation of 3D models of long bones from healthy human volunteers. While bones do not generate a useful signal in clinical MRI due to extremely short transverse relaxation times, the bone geometry can be delineated using the signal generated from the surrounding soft tissue. Even though this has been demonstrated in some studies, the accuracy of these models has to be quantified using in vitro and in vivo studies before such models can be used for implant design. This chapter discusses the investigation carried out to quantify the accuracy of 3D models reconstructed using a currently available clinical MRI scanner. Section 5.2 discusses the relevant literature, and Section 5.4 briefly introduces the methods used for the study. More details of the study are available in the published journal article that is presented at the end of this chapter. Basic principles of the MRI scanner and their relevance to bone imaging have been discussed in Chapter 2. The segmentation techniques and 3D modelling methods used in this study have been previously investigated and validated as a part of the PhD project and the details are available in Chapter 3.

67

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

5.2 Imaging of skeletal system with MRI MRI is designed to scan soft tissues utilising 1H nuclei as the source of signal. The signal intensity generated from a particular tissue type in MRI is determined by the longitudinal relaxation time (T1), transverse relaxation time (T2), and proton density of the tissue type. Different soft tissues of human or other mammalian bodies have different T1 and T2 values. Hence, by using different TR and TE values, a better contrast between two different soft tissue types can be obtained. Due to the superior contrast obtained, MRI has become the method of choice for quantitative studies of cartilage, muscles and other soft tissues [1, 13-17]. In contrast to the soft tissues, cortical bone (including ligaments and menisci) has extremely short transverse relaxation times (T2) and does not produce an adequate signal in clinically used pulse sequences [10-12]. Hence, the visualisation of the bone structure is not possible with clinically used pulse sequences. This might be achieved with special pulse sequences that have ultra-short TE values, in which TE is reduced to 0.07-0.20 ms from usual values of 4-10 ms [10, 12]. Human or other mammalian long bones are surrounded by a good bulk of muscles and other soft tissues. These soft tissues are capable of producing MR signals with high intensity when clinically used pulse sequences are employed and, hence, produce a high contrast between the cortical bone and surrounding soft tissues. Using this high contrast, it is possible to identify the cortical bone geometry in MRI images acquired with the clinically used protocols (Figure 5.1).

68

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

Figure 5.1: Cross sections of CT (left) and MRI (right) from the same anatomical location of a sample (In the CT image, the cortical bone appears in high intensity and can be clearly identified from the surrounding soft tissue. In the MRI image, cortical bone appears in black as it does not generate a significantly high signal; however, the outer cortex can be identified due to the signal generated by surrounding soft tissues) MRI has long been used for bone imaging, mainly for diagnosing metastatic disease, for computer assisted surgery (CAS) and for bone motion kinematic studies. The skeletal system is one of the main targets for cancer metastases and MRI has been a superior imaging method to detect these metastases over the other imaging techniques (CT and plain x-ray) [18]. MRI has also been used for quantification of the trabecular bone structure in several studies [119-121]. The next main use of MRI related to bone imaging is for CAS of the spine [19, 20, 24]. The usual practice for CAS of the spine is to generate 3D models of the spine using CT to help the accurate placement of pedicle screws. Due to the radiation exposure of CT, Hoad et al. [24] have developed an MRI imaging sequence that can be used for generating 3D models of the spine. A double echo sequence was used and a 3D model of the vertebrae was generated by manually segmenting the image data. The model was compared with a similarly generated model using CT. The results show that the accuracy of the MRI based model is 90% compared to the 100% accuracy of the CT based model. Bone kinematic studies have also been a major research area that has used MRI in place of gold standard CT due to the high radiation exposure [21, 25, 122-124]. CT has been the typical image acquisition method for quantification of position of bones during various movements due to the very short image acquisition time and the high soft tissue bone contrast. Wolf et al. [25] imaged the feet of five volunteers with a 3T MRI system (resolution 0.39 mm × 0.39 mm × 0.7 mm) using a 3D T 1 weighted 69

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

gradient echo sequence. The metatarsal bones were segmented using an intensity thresholding segmentation method. Comparison of the MRI based 3D models with a reference standard was not conducted; however, Wolf et al. recommends use of MRI for foot bone motion quantification. Fassbind et al. [21] used a 1.5T MRI scanner to quantify foot bone motion and obtained kinematic characteristics similar to those cited in other published studies that used CT. Pillai et al. [123] studied the wrist bone motion using 3D models of radius, scaphoid and lunate generated from 1.5T MRI scanner for which a low resolution (0.31 mm × 0.31 mm × 2 mm ) 3D FLASH sequence was used. Manual segmentation was performed excluding the cartilage from bone. The kinematic analysis showed results similar to those published. In addition to those mentioned above, a number of studies have been conducted to investigate the kinematics of the tibio-femoral joint using MRI. DeFrate et al. [125] and Chen et al. [126] studied the knee kinematics generating 3D models of the distal part of the femur and the proximal part of the tibia using MRI, while Hao et al. [127] used MRI to generate a finite element model of a knee joint. All three studies obtained accurate results for kinematic studies. Even though all of the studies described above have reconstructed the 3D models of various parts of long bones and small bones using MRI, a validation with a proper reference standard was not employed to quantify the accuracy of those models. Musculoskeletal models that represent bone including cartilage, ligaments and muscles have been successfully generated using a combination of MRI and CT. Lee et al. [1] conducted a study using five porcine femora to generate a combined MRI and CT model in which MRI was used to reproduce soft tissues, while CT was used for bones. The in plane resolution of CT and MRI data was 0.4 mm × 0.4 mm and 0.3 mm × 0.3 mm respectively. CT had slice reconstruction interval of 0.625 mm, while the slice thickness of MRI was 1.2 mm. The 3D models were reconstructed by manually segmenting the image data. The 3D models derived from MRI were registered to the CT models with a surface matching accuracy of 0.7 ± 0.1 mm. Moro-oka et al. [124] conducted a study to compare three-dimensional kinematic measurements from single plane radiographic projections. Three knee joints of human volunteers were scanned using CT and MRI scanners with the resolution of 0.35 mm ×0.35 mm × 1.00 mm and 0.39 mm ×0.39 mm ×1.00 mm respectively. 3D 70

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

models of the knee joint were reconstructed using a commercial software package; however, the method used for the segmentation has not been mentioned. The surface matching of the MRI models to CT models presented differences of -0.11 ± 0.81 mm, -0.23 ± 0.48 mm and -0.12 ± 0.60 mm for femora of three subjects, and -0.14 ± 0.67 mm, -0.13 ± 0.48 mm and -0.15 ± 0.77 mm for tibiae. The studies mentioned above have shown that a sub voxel level accuracy can be obtained for 3D models of bones using MRI. The main drawback of the studies is that MRI models have not been validated using a proper reference standard. Most of the studies have not used MRI to generate the complete 3D models of long bones which are necessary for the design of trauma fixation plates and nails. The studies have also not focused on quantifying the accuracy of the medullary canal of a long bone that is important for designing intramedullary nails. Therefore, studies that focus on quantifying the geometric accuracy of complete 3D models of long bones and the medullary canal are required to inform the improved design of orthopaedic implants using 3D models based on MRI.

5.3 Aims of the study The aim of the study was comprehensive quantification of the accuracy of 3D models based on MRI compared to the 3D models based on CT, and their formal validation using a reference standard based on the contact surface scanner.

5.4 Methods MRI and CT scans of five intact femora were obtained by scanning five intact ovine hind limbs. Ovine femora were used, as human volunteers cannot be CT imaged due to radiation exposure, and contact mechanical scanning cannot be used for validation. The sample size calculation showed that the required sample size to detect a difference of 0.08 mm with standard deviation of 0.02 is four. A 1.5T MRI scanner (Siemens Magnetom Avanto) and a 64 slice CT scanner (Phillips Brilliance 64) were employed for scanning of the limbs. The MRI scanner was used with a 3D flash sequence, TR = 11 ms, TE = 4.94 ms, FA = 15º and 0.45 mm ×0.45 mm resolution with 1 mm slice thickness. These parameters were chosen as they produced the best results for different parameter combinations used in the pilot study. The CT scanner was used with a 0.4 mm ×0.4 mm in plane resolution and 0.5 mm slice spacing, kVp 71

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

= 140 and mAs = 231. The segmentation of MRI and CT data was performed using the multi-threshold segmentation method combined with the threshold selection method developed as a part of this research project (See Chapter 3). The triangular meshed contact scanner and microCT generated surfaces of the soft tissue free bones were used as reference standards for outer and inner surfaces respectively. The method of generating the reference standard has been described in Section 3.8.3. Using the point to point comparison method described in Section 4.3, comparisons were carried out between the MRI based models and the reference models, the CT based models and the reference models, and the MRI based models and the CT based models. A detailed description of the methods is presented in the journal article presented at the end of this chapter.

5.5 Results Comparison of the MRI based and CT based 3D models to the reference models showed average errors of 0.23 mm and 0.15 mm respectively. Statistically, there was no significant difference between the 3D models based on two methods (p = 0.067). A detailed results section is available in the publication presented at the end of this chapter.

5.6 Summary, discussion and conclusion MRI has shown to be an ionising radiation free potential alternative for CT. A number of kinematic studies, finite element studies and studies of the diagnosis of metastatic disease have successfully used MRI as an alternative to CT for scanning the skeletal system, even though these studies have not validated MRI based 3D models with a proper reference standard. Few studies [1, 124] surface matched MRI based 3D models to CT based models and reported sub voxel level accuracies; however, a reference standard such as a laser or contact scanner has not been used for the validation. The present study aimed at quantifying the accuracy of the surface geometry of MRI based 3D models and the CT based 3D models, using state of the art dense triangular meshed surface scans of the outer and inner surfaces of femora as the reference standard. The study acquired MRI and CT data from five sheep femora with intact soft tissues and intact joints. 3D models were generated using a multilevel 72

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

thresholding method in combination with a method to select the threshold level for the particular region. The resulting accuracy of the MRI based 3D models (average deviation = 0.23 mm) was comparable to that of CT based models (average deviation = 0.15 mm). There was no statistically significant difference between the two methods. This indicates that the 3D models based on MRI can be used as an alternative to CT for 3D reconstruction of long bones. The statistical analysis also shows that to detect this difference (0.08 mm) a sample size of four is sufficient. The diaphyseal region of the femora presented an accuracy of 0.15 mm, while the proximal and distal regions which have complex geometric shapes gave a relatively lower accuracy. The poor contrast obtained from the 1.5T MRI scanner for these articular regions forced to manually segment most of these regions, potentially introducing errors to the final 3D surfaces. The long scanning time of the MRI compared to the CT scanning time poses a number of additional limitations when MRI is used for scanning of human volunteers. The motion artefact from random movements is one of the important limitations and this is addressed in Chapter 7 of this thesis. The longer segmentation time of the MRI images compared to the segmentation time of the CT images also limits the use of MRI for imaging of long bones. This study used ovine femora as the study sample. In order to apply these methods to the much larger human long bones, additional studies using human long bones are desirable before applying these methods on humans. One limitation of the study is that only two of five bones were used for the reconstruction of the inner medullary canal due to the inadequate contrast in other bones. The fat/water only imaging would be advantageous here and would be suggested as possibility in future. The study showed that MRI can generate 3D models of long bones with accuracy comparable to that of CT models. Using a higher field strength scanner, typically 3T, the current drawbacks of poor contrast in certain anatomical regions and the longer scanning times can be potentially overcome. The next chapter will investigate the use of 3T MRI to overcome these drawbacks using human volunteers as the study samples.

73

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones

5.7 Paper 3: Quantification of the accuracy of MRI generated 3D models of long bones compared to CT generated 3D models (in press)

74

Chapter 5: Magnetic resonance imaging for 3D reconstruction of long bones Medical Engineering & Physics xxx (2011) xxx-xxx

Contents lists available at Scie nceD irect

Medical Engineering & Physics journal homepage: www . elsev ie r .co m /loca t e/medengp h y

ELSEVIER

Quantification of the accuracy of MRI generated 30 models of long bones compared to CT generated 30 models Kanchana Rathnayaka a. Konstantin I. Momot b, Hansrudi Noserc, AndrewVoJp d, Michael A. Schuetz a,d, Tony Sahama b, Beat Schmutza,. • lnstirute of Health and Biomedical Innovation, 60 Musk Avenue, Kelvin Grove, QLD 4059, Australia b Queensland University of Technology, 2 George Street, Brisbane, QLD 4000, Australia ' AO Research Institute Davos. Clavadelerstrasse 8. 7270 Davos. Switzerland d Princess Alexandra Hospital. 199 1pswich Road. Woolloongabba. Brisbane. QLD 4102, Australia

ARTICLE

INFO

Due to copyright restrictions, this article is not available here. Please consult the hardcopy thesis available from QUT Library or view the published ABSTRACT version online at:

Article history: Received 18 February 2011 Received in revised form 25 July 2011 Accepted 27 July 2011 Keywords:

MRI CT 3D models Femur

Orthopaedic fracture fi xation implants are increasingly being designed using accurate 3D models of long bones based o n computer tomography (Cf). Unlike cr. magnetic resonance imaging (MRI) does http://dx.doi.org/10.1016/j.medengphy.2011.07.027 not involve ionising radiation and is therefore a desirable alternative to cr. This study aims to quantify the accuracy of MRI-based 3D models compared to er-based 3D models of long bones. The femora of five intact cadaver ovine limbs were scanned using a 1.5 T MRI and a er scanner. Image segmentation of er and MRI data was performed using a multi-threshold segmentation method. Reference models were generated by digitising the bone surfaces free of soft tissue with a mechanical contact scanner. The MRIand er-derived models were validated against the reference models. The results demonstrated that the er-based models contained an average error of0.15 mm w hile the MRI-based models contained an average error of 0.23 mm. Statistical validation shows that there are no significant differences between 3D mode ls based on er and MRI data. These results indicate that the geometric accuracy of MRI based 3D models was comparable to that ofCf-based models and therefore MRI is a potential alternative to er for generation of 3D models with high geometric accuracy. © 2011 IPEM. Published by Elsevier Ltd. All rights reserved.

1. Introduction Three-dimensional (3D) models of long bones with a high geometric accuracy are widely utilised by medical engineering research and in clinical practice; the design of orthopaedic fracture fixation implants [1 ,2]. computer aided surgery simulations [3,4] and fracture healing models [5,6] are just a few examples. Computed tomography (CT) has become the gold standard for scanning of bones to produce 3D models with high geometric accuracy. Due to high radiation exposure, CT cannot be used to scan healthy human volunteers. Therefore, an alternative method for the scanning of long bones of the healthy human population needs to be investigated. Among various uses of 3D models, orthopaedic implant design particularly requires 3D models with high geometric accuracy to produce implants with a better fit to the patients' anatomy [ 1,3 ]. Furthermore, the anatomically pre-shaped implants are often designed based on the Caucasian population and thus the size and

*Corresponding author. Tel.: +61 7 3138 6238; Fax: +61 7 3138 6030. f -mail address: [email protected] (B. Schmutz).

shape do not accurately match the Asian population. Therefore, those pre-shaped implants still need some optimisation for a better anatomical fit to people of different ethnic origins and age groups [2]. Ethnicity and age are two important factors which determine the shape and size of bones [7,8]. Thus, a database with accurate bone data from different ethnic and age groups is essential for this purpose. Some institutions have already started developing such databases using CT imaging of cadaver bones [9] but the majority of these bones are from older donors (>60 years) therefore do not represent the young patient population. Furthermore. cadaver bones can seldom be chosen according to the researchers' need (e.g. gender or specific subject height) due to the limited availability. Therefore, there is a need to collect bone data from healthy human volunteers who represent that part of the patient population for whom no CT data exists or can be acquired. This would facilitate researchers' access to specific population groups for the purpose of obtaining high-quality anatomical image data. CT scanning of healthy human volunteers is not ethically justifiable due to the high radiation exposure [10,11 ]. Studies investigating CT imaging protocols that use low radiation doses, while keeping the original image quality, have become an important part

1350-4533/$- see front matter © 2011 IPEM. Published by Elsevier Ltd. All rights reserved. doi: 10.1 016/j.medengphy.2011.07.027

75

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Chapter 6 Higher field strength MRI scanning of long bones for generation of 3D models 6.1 Introduction As discussed in Chapter 5, 1.5T MRI offered acceptable accuracy for reconstruction of 3D models from long bones. However, the images of long bones acquired at 1.5T MRI need to be further improved to overcome limitations such as poor contrast in articular regions and long scanning times. The high field strength scanners are promising to offer higher signal to noise ratio (SNR) levels [22] which can potentially be used to overcome these limitations of 1.5T scanners. The higher SNR obtained at higher field strengths could, in principle, be used either for higher resolutions or for higher contrast levels. In this study, the signal gain was investigated in the form of contrast to noise ratio (CNR). As the intrinsic SNR of an MRI system is approximately proportional to the main magnetic field (B0), if all the parameters, subjects and radio frequency (RF) coils are equivalent, scanners with 3T magnets should theoretically yield approximately double the SNR at 1.5T [128]. However, the increased main magnetic field affects the tissue parameters such as the longitudinal relaxation time (T 1) and transverse relaxation time (T2). Therefore, before the use of 3T scanners for musculoskeletal imaging, it is necessary to investigate the effect of the higher field strength on the tissue parameters (e.g. T1 or T2) and imaging artefacts, and to optimise the imaging protocols accordingly. Section 6.2 of this chapter discusses the theoretical increase of SNR at 3T. Relevant literature on the use of the 3T MRI system for scanning of the musculoskeletal system is discussed in Section 6.3 . Section 6.5.1 discusses the basic principles of the methods used for quantification of image quality of an MRI system. While a detailed 83

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

description of the methodology used in this study is available in the publication, a brief introduction is included in Sections 6.5.2 and 6.5.3.

6.2 Theoretical consideration of increased SNR at 3T The theoretical signal gain in an MRI system is proportional to the square of the main magnetic field; thus, the signal gain at 3T should be 4 times that at 1.5T, as given in the equation below:

S

B0

2

6.1

Where S = signal, B0 = main magnetic field,

= gyromagnetic ratio.

The noise level (N) of a MRI system is proportional to the Larmor frequency and, hence, to the main magnetic field. Thus, the noise level at 1.5T becomes two fold at 3T (6.2). Therefore, the actual SNR gain at 3T is two times that of 1.5T.

N

SNR

v0

N 4S 2N

B0

2

6.2

6.3

Where N = Noise level, v0 = Larmor frequency, B0 = main magnetic field, SNR = signal to noise ratio and S = signal [23].

6.3 3T MRI for musculoskeletal system imaging Scanners with higher field strengths, typically 3T, became available for clinical scanning in the 1990s. Since then, a number of quantitative and qualitative comparisons between 1.5T and 3T have been carried out, mainly to compare various soft tissue compartments [43, 129-134]. In comparison, fewer articles have been published comparing 3T imaging of the musculoskeletal system. Of these, some are related to quantifying the cartilage morphology [135, 136] and the spin relaxation times or anatomical structure demonstrations [128, 137-139]. A relatively large number of review articles have been published by MRI experts regarding various aspects of 3T or high field MRI [23, 140-146].

84

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Among the articles published regarding musculoskeletal imaging with 3T, Stehling et al. [137] assessed a multi-contrast high resolution imaging protocol for imaging of the wrist of 10 volunteers using 3T and 1.5T scanners. The imaging protocol at 3T had half the in plane resolution and half the slice thickness of those at 1.5T imaging protocol (0.5 mm × 0.5 mm and 3.0 mm respectively). The idea was to use the SNR gain at 3T for better image resolution. The qualitative assessment showed that the structure and overall image quality was significantly higher in 3T (p 0.5 R1; S0 > 0.5 S0; R1 < 0; S0 < 0. The average R1 values in the „muscle‟ and „bone marrow‟ compartments were then determined by averaging the fitted R1 values over the „non-rejected‟ voxels within the appropriate ROI (~1000 voxels for the muscle and ~300 voxels for the marrow). For the measurement of T2*, the RF excitation pulse was set to

RF

= 15o, the

repetition time to 16 ms, the number of averages to 1, and a series of gradient TE values were used. At 1.5T, the TE values used were 4, 5, 6, 8, 10 and 12 ms. At 3T, 102

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

the TE values used were 4, 5, 7, 9, 12 and 14.8 ms. For each voxel within the image, the T2 value was determined from a two-parameter nonlinear least-squares fit of the intensity as a function of TE:

S (TE ) S0 e

TE R2*

(2)

The fitted value of R2* was taken as the apparent transverse relaxation rate, 1/T2*, in the respective voxel. Fit quality control and T2* averaging over muscle and marrow were performed as described above for the T1. T1 and T2 * processing was performed using custom-written Mathematica (Wolfram Research, Champaign, IL, USA) code running on a desktop PC. SNR was calculated for the „muscle‟ and „bone marrow‟ tissue types of image series. ROIs were selected at five sites of an image slice of each of the image stack, as indicated in Figure 4-b. The SNR was calculated using the method described in the next section. SNR and CNR for comparison of MR images SNR and CNR were used to compare image quality between images obtained from 1.5T and 3T scanners. CNR is one of the most important parameters as the contrast between bone and the soft tissue is the key feature that is responsible for an accurate segmentation of the bone. In the basic form, SNR is the ratio of a signal to the background noise, while CNR is the ratio of contrast to the background noise (Equations 3 & 4).

SNR

Signal Noise

(3)

CNR

Signal1 Signal2 Noise

(4)

When the equations were applied to the MR images, the mean intensity of the specific tissue type was considered as the signal and the standard deviation of the background was considered as the noise level [19]. The noise statistics derived correction factor

2 /(4

) [20] was introduced to standardise the SNR and CNR 103

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

values derived. In this study, the background noise level was not measured due to the unevenly distributed noise in background and thus the noise level of the cortical bone was used to calculate the SNR and CNR [20]. Thus, with the noise statistics derived factors, the equations used to calculate the SNR and CNR of MR images were as follows: SNR 4

CNR

M tissue 2 STDbone

M tissue1 M tissue 2 2 STDbone 4

(5)

(6)

Where, SNR = signal to noise ratio, CNR = contrast to noise ratio, M tissue = Mean intensity of the tissue and STDbone = standard deviation of mean intensity of cortical bone. Comparison of the images obtained from 1.5T with 3T MRI SNR and CNR measurements were taken in the proximal articular, proximal diaphyseal, mid diaphyseal, distal diaphyseal and distal articular regions of the femur and of the tibia. In diaphyseal regions (Figure 2), SNR and CNR were measured (Figure 3) at five sites around the bone (Figure 4-b) in axial image slices. In articular regions, a varying number of sites were used (Figure 4-a, c, d, e & f) and coronal sections were used, with the exception of the distal articular region of the femur for which axial images were used. The measurements were taken in three consecutive image slices at any given site.

104

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Figure 2. The diaphyseal regions of femur (top) and tibia (bottom) where the axial image slices were obtained for the calculation of SNR and CNR.

Figure 3: In each site of the diaphyseal regions, pixel samples were obtained from bone marrow, cortical bone, and Muscle.

105

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Figure 4. ROIs selected at four/five positions in each tissue type in: a-femoral head, b- mid femoral diaphysis, c- distal femoral diaphysis, d- distal femoral articular, eproximal tibial articular, and f- distal tibial articular regions as shown in the figure (left and right images are from two different planes). SNR and CNR were measured in muscle, bone and bone marrow tissue types at diaphyseal regions of both femora and tibiae, with the exception of the distal diaphyseal region of the femora where the bone was surrounded by other soft tissues in addition to the muscle tissue (mainly fat and fibrous tissue). In this region the measurements were taken in these soft tissues in addition to muscle. In the articular regions, the bone does not come into contact with the muscle tissue but with various other tissues such as fat, tendons, fibrous capsules and synovial fluid. Moreover, the articular regions no longer contain bone marrow, and the medulla is basically composed of a mixture of trabecular bone and bone marrow. Thus, the measurements were taken in soft tissues, bone and the medulla. Statistical differences of SNR and CNR values between the 1.5T and 3T images were calculated using one way ANOVA. The level of statistical significance was set to p ≤ 0.05. The validation was performed using PASW Statistics 18 software package.

106

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Results The measurement of the longitudinal relaxation time (T1) and the apparent transverse relaxation time (T2*) was carried out using a series of images acquired with varying TR and varying TE values, respectively. The measured T1 value of the muscle was 1.5 ± 0.2 s at 3T and 0.9 ± 0.1 s at 1.5T. The measured T1 values of the voxels in the bone marrow compartment were 0.25 ± 0.03 s at 1.5T and 0.30 ± 0.07 s at 3T. The apparent transverse relaxation time, T2*, of the muscle was measured as 0.029 ± 0.007 s at both 1.5T and 3T. The T2* of the bone marrow could not be measured reliably. The SNR calculation of the images obtained with varying TR, TE and FA values (Figure 5) showed the following trends. SNR of muscle and bone marrow increased with the TR while SNR of muscle and bone marrow declined with the TE in both 1.5T and 3T filed strengths; and SNR of muscle had downward trend with FA, while SNR of bone marrow had upward trend with the FA in both 1.5T and 3T field strengths.

Figure 5. Change of SNR with varying TR, TE and FA at 1.5T and 3T. 107

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

The comparison between 1.5T and 3T images of the femora produced the following results (Figure 6). In the mid diaphyseal region 3T had the highest CNR and SNR for muscles (CNR = 4.49, 7.29 and SNR = 7.50, 10.00 for 1.5T and 3T respectively) and 1.5T had the highest CNR and SNR for bone marrow (CNR = 6.49, 5.70 and SNR = 9.66, 8.67 respectively for 1.5T and 3T). In the proximal diaphyseal region, CNR and SNR of muscle at 3T were slightly higher than 1.5T and CNR and SNR of bone marrow was higher at 1.5T. In the distal diaphyseal region, CNR and SNR of the other soft tissues were slightly higher at 3T (CNR = 4.74, SNR = 6.97) than 1.5T (CNR = 4.54, SNR = 6.96); however, CNR and SNR for muscles were slightly higher at 1.5T compared to 3T. For the same region, CNR and SNR of medulla were higher in 1.5T compared to 3T.

Figure 6: CNR and SNR of diaphyseal regions of femur (TR = 11 ms and TE = 4.66 ms at both 1.5T and 3T, * = statistically significant). 108

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

SNR and CNR measurements of four sites at the proximal articular region and five sites at the distal articular region show that 3T MRI gives higher SNR and CNR for all the regions with the exception of region -4 of the distal articular region that has higher SNR and CNR for 1.5T (Figure 7). Images illustrating the improvement in image contrast are shown in Figure 8.

Figure 7. Proximal and distal articular regions of the femur (TR = 11 ms and TE = 4.66 ms at both 1.5T and 3T, * = statistically significant).

109

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Figure 8. Comparison of 1.5T images to 3T images of the proximal region (top) and the mid shaft (bottom) of the femur (TR = 11 ms and TE = 4.66 ms at both 1.5T and 3T). In tibia, the proximal diaphyseal region, muscles presented higher SNR and CNR values for 3T MR images while medulla showed similar SNR and CNR values for both 1.5T and 3T. For the mid diaphyseal region; however, 1.5T showed higher SNR and CNR than 3T (SNR = 15.4 and 14.5, CNR = 13.3 and 12.4 respectively for 1.5T and 3T). For the distal diaphyseal region higher CNR and SNR was reported for 3T (Figure 9).

110

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Figure 9. CNR and SNR values of diaphyseal regions of tibia femur (TR = 11 ms and TE = 4.66 ms at both 1.5T and 3T, * = statistically significant). CNR and SNR measured at four sites in both the proximal articular region and the distal articular region of tibiae showed higher CNR and SNR for 3T images with the exception of the region -3 of the distal articular region (Figure 10).

111

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

Figure 10. CNR and SNR of articular regions of tibia (TR = 11 ms and TE = 4.66 ms at both 1.5T and 3T, * = statistically significant).

Discussion The study aimed to quantitatively compare the MR image quality at two applied magnetic field strengths, 1.5T and 3T, using the femora and tibiae of five healthy volunteers as the study sample and SNR and CNR as the comparison parameters. An investigation was also carried out to optimise the imaging protocol at 3T by identifying the optimum TR, TE and FA values at that field strength. The effect of the magnetic field on the T1 and T2 of the tissues imaged (muscle and bone marrow) was also investigated. The T1 of the muscle was strongly dependent on the applied magnetic field strength (B0): The T1 at 3T (1.5 s) was more than 50% longer than that at 1.5T (0.9 s). The apparent T1 values of the bone marrow exhibited significantly weaker field dependence, with the apparent T1 values at the two fields differing by ~15% (0.25 s at 1.5T and 0.30 s at 3T). (We use the term „apparent T1‟ for the bone marrow because no fat suppression was used, and the measured T1 in this tissue can therefore include contributions from both lipid and water.) The lengthening of the T1 values of 112

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

both tissues with the increasing B0 is consistent with the well-established body of knowledge concerning the relaxometry of biological tissues [18, 23, 24]. It is also consistent with the fact that longitudinal relaxation is controlled by fast molecular motions[25]; that is, motions whose time scale is comparable to the Larmor precession frequency of the MRI systems used in this study (~10 ns). The relatively small increase of the apparent T1 of bone marrow can be attributed to the relatively low mobilities of lipid molecules and water molecules in a lipid-rich environment. This observation is consistent with the field dependence of the T1s of lipid and water protons previously observed in a model lipid/water system [26]. The apparent transverse relaxation time, T2*, of the muscle exhibited no discernible dependence on the applied magnetic field. This observation can be rationalised as follows. T2* is a complicated function dependent on the local inhomogeneities of the static magnetic field, slow molecular motions, fast molecular motions, and chemical exchange between „free‟ and „bound‟ states of water molecules. (The last three factors determine the true transverse relaxation time, T2.) The four factors listed serve to shorten, shorten, lengthen, and shorten T2* with the increasing B0, respectively [27]. The true T2 in muscle has been reported variously to become slightly shorter [19] or slightly longer [18] with the increasing B0. Under the conditions of the present study, the effects of the four factors listed evidently nearly cancel each other out, resulting in the absence of a significant field dependence of T2*. When a 3D FLASH sequence was used, the SNR of both muscle and bone marrow increased upon increasing TR from 10 ms to 50 ms. Beyond 50 ms, SNR at 3T started to decline and TR > 50 ms was not used with 1.5T imaging due to practical difficulties of setting up the scanner with TR = 100 ms. Even though a higher SNR can be obtained at higher TR, doubling TR in turn doubles the scanning time. Compared to the scanning time of 65 minutes to scan the complete lower limb with TR = 11, TR beyond this would result in extremely long scanning times that are impracticable in the clinical environment, but also due to the increased risk of motion artefacts resulting from long scanning times. With increased TE, SNR dropped in both muscle and bone marrow tissues at 1.5T and in muscle tissue at 3T; however, SNR increased in bone marrow at 3T (Figure 113

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

5). FA presented converse SNR results for muscle and bone marrow. At both 1.5T and 3T, SNR of muscles declined on increasing FA whereas SNR of bone-marrow increased. Based on the results obtained, the protocol used to scan human volunteers was determined to have TR = 11, TE = 4.66 and FA = 7 for both 1.5T and 3T scanners. Comparison of images obtained at 1.5T to 3T showed that, in general, 3T MRI generates images with a high contrast between bone-muscle and bone-soft tissue interfaces. The mid diaphyseal region of the femur, and the proximal and distal diaphyseal regions of the tibia, presented a greater increase in CNR and SNR in the bone-muscle interface, while the proximal diaphyseal region of the femur showed slight increase. Among them, the mid diaphyseal region of the femur showed statistically significant increase in SNR and CNR at 3T. The distal diaphyseal region of the femur and the mid diaphyseal region of the tibia did not show any increase in CNR or SNR for muscle at 3T, the reason for this could not be determined. However, there was a slight increase in CNR and SNR at soft tissue-bone interface of the distal diaphyseal region of the femur. The reason why CNR and SNR were lower in these regions could not be determined. CNR at the bone marrow-bone interface was higher at 1.5T than 3T in all the cases and this was statistically significant in mid diaphyseal region of tibia. As mentioned at the beginning of the discussion, T1 of bone marrow (0.25 ± 0.03 s at 1.5T and 0.30 ±0.07 s at 3T) is comparatively shorter than T1 of muscle tissue (0.86±0.14 s and 1.5±0.15 s at 3T). As the extremely short TR value (11 ms) have been used for both 1.5T and 3T scanning, tissues with longer T1 (muscle in this case) produce a low signal due to inadequate recovery of the transverse component of the net magnetisation vector. This is the main reason why bone marrow produced a higher signal compared to the muscle. The low CNR of bone-bone marrow interface at 3T is unlikely to affect the segmentation process as the obtained CNR is sufficient for an accurate segmentation of the medullary canal. Compared to the outer cortex, the inner cortex has a relatively simple, bone-bone marrow interface in the medullary canal. Articular regions of both the femur and tibia showed increased CNR and SNR for 3T, with the exception of one site in the distal articular region of the femur and the 114

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

second site at the distal articular region of the tibia. These differences were statistically significant in two regions in each of the proximal articular region of femur, proximal articular region of tibia and distal articular region of tibia for CNR. In distal articular region tibia, the differences were also statistically significant for SNR at two sites. The reason for this difference in CNR and SNR could be due to the number of different interfaces present at the articular regions (bone-ligament, bonetendon, bone-synovial fluid, bone-synovial membrane and bone-cartilage). These different tissue types exhibit different MRI properties (T1, T2* and proton density) that result in various contrast levels at the articular regions. This increase the partial volume effect at articular regions and therefore only the average CNR and SNR can be measured in these regions. However, increased CNR at most of the sites of the articular regions will potentially facilitate the segmentation process by improving the accuracy, which was a problem in 1.5T MR imaging of those regions. Overall, 3T MRI generated images with higher quality for most of the anatomical regions of the femur and tibia. Even though the theoretical doubling of SNR gain is not achievable due to the practical reasons, the articular regions had impressively higher CNR and SNR values. These are the regions where segmentation at 1.5T was difficult and this increased CNR is expected to significantly facilitate the segmentation process of the articular regions [6, 7]. CNR and SNR of distal femur and mid diaphysis of tibia were not improved; however, these regions could be segmented accurately with 1.5T images [6]. At the same time, the obtained higher contrast levels at bone-muscle and bone–bone marrow interfaces will potentially improve the accuracy of segmentation and in addition decrease the time required for the segmentation. This study investigated the improvement of the image contrast by using higher field strength MRI. Another important aspect that needs to be improved through future studies is the scanning time, which is considerably longer compared to CT at present.

Acknowledgement This research was supported in part by Synthes GmbH. The last author has received an industrial scholarship from Synthes GmbH.

115

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

References [1].

Schmutz B, Rathnayaka K, Wullschleger ME, Meek J, Schuetz MA. Quantitative fit assessment of tibial nail designs using 3D computer modelling. Injury 2010;41(2):216-219.

[2].

Schmutz B, Wullschleger ME, Kim H, Noser H, Schuetz MA. Fit Assessment of Anatomic Plates for the Distal Medial Tibia. Journal of Orthopaedic Trauma 2008;22(4):258-263.

[3].

Schmutz B, Wullschleger ME, Noser H, Barry M, Meek J, Schuetz MA. Fit optimisation of a distal medial tibia plate. Computer Methods in Biomechanics & Biomedical Engineering 2011;14(4):359-364.

[4].

Park A, DiStefano J, Nguyen T, Congruency of Scapula Locking Plates Implications for Implant Design, in 56th Annual Meeting of the Orthopaedic Research Society 2010: New Orleans, LA. p. Poster No. 1746.

[5].

Kozic N, Weber S, Büchler P, Lutz C, Reimers N, González Ballester MÁ, Reyes M. Optimisation of orthopaedic implant design using statistical shape space analysis based on level sets. Medical Image Analysis 2010;14(3):265275.

[6].

Rathnayaka K, Momot KI, Noser H, Volp A, Schuetz M, Sahama T, Schmutz B. Quantification of the accuracy of MRI generated 3D models of long bones compared to CT generated 3D models. Medical Engineering & Physics 2011(in press DOI: 10.1016/j.medengphy.2011.07.027).

[7].

Schmutz B, Volp A, Momot K, Pearcy M, Schuetz M. Using MRI for the imaging of long bones: First Experience. Journal of Biomechanics 2008;41(Supplement1):S188.

[8].

Edelstein WA, Glover GH, Hardy CJ, Redington RW. The intrinsic signal-tonoise

ratio

in NMR

imaging.

Magnetic

Resonance

in Medicine

1986;3(4):604-618. [9].

Soher BJ, Dale BM, Merkle EM. A Review of MR Physics: 3T versus 1.5T. Magnetic Resonance Imaging Clinics of North America 2007;15(3):277-290.

116

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

[10].

Rakow-Penner R, Daniel B, Yu H, Sawyer-Glover A, Glover GH. Relaxation times of breast tissue at 1.5T and 3T measured using IDEAL. Journal of Magnetic Resonance Imaging 2006;23(1):87-91.

[11].

Tsurusaki M, Semelka RC, Zapparoli M, Elias Jr J, Altun E, Pamuklar E, Sugimura K. Quantitative and qualitative comparison of 3.0 T and 1.5 T MR imaging of the liver in patients with diffuse parenchymal liver disease. European Journal of Radiology 2009;72(2):314-320.

[12].

Ramalho M, Altun E, Herédia V, Zapparoli M, Semelka R. Liver MR Imaging: 1.5T versus 3T. Magnetic Resonance Imaging Clinics of North America 2007;15(3):321-347.

[13].

Frayne R, Goodyear BG, Dickhoff P, Lauzon ML, Sevick RJ. Magnetic resonance imaging at 3.0 Tesla: Challenges and advantages in clinical neurological imaging. Investigative Radiology 2003;38(7):385-402.

[14].

Beyersdorff D, Taymoorian K, Knosel T, Schnorr D, Felix R, Hamm B, Bruhn H. MRI of Prostate Cancer at 1.5 and 3.0 T: Comparison of Image Quality in Tumor Detection and Staging. American Journal of Roentgenology 2005;185(5):1214-1220.

[15].

Stehling C, Langer M, Bachmann R, Kraemer S, Kooijman H, Heindel W, Vieth V. Three-Tesla Magnetic Resonance Imaging of the Wrist: Diagnostic Performance Compared to 1.5-T. Journal of Computer Assisted Tomography 2009;33(6):934-939.

[16].

Duewell SH, Ceckler TL, Ong K, Wen H, Jaffer FA, Chesnick SA, Balaban RS. Musculoskeletal MR imaging at 4 T and at 1.5 T: comparison of relaxation times and image contrast. Radiology 1995;196(2):551-555.

[17].

Gold GE, Suh B, Sawyer-Glover A, Beaulieu C. Musculoskeletal MRI at 3.0 T: Initial Clinical Experience. American Journal of Roentgenology 2004;183(5):1479-1486.

117

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

[18].

Stanisz GJ, Odrobina EE, Pun J, Escaravage M, Graham SJ, Bronskill MJ, Henkelman RM. T1, T2 relaxation and magnetization transfer in tissue at 3T. Magnetic Resonance in Medicine 2005;54(3):507-512.

[19].

Kuo R, Panchal M, Tanenbaum L, Crues JV. 3.0 Tesla imaging of the musculoskeletal

system.

Journal

of

Magnetic

Resonance

Imaging

2007;25(2):245-261. [20].

Dietrich O, Raya JG, Reeder SB, Reiser MF, Schoenberg SO. Measurement of signal-to-noise ratios in MR images: Influence of multichannel coils, parallel imaging, and reconstruction filters. Journal of Magnetic Resonance Imaging 2007;26(2):375-385.

[21].

Chang G, Friedrich KM, Wang L, Vieira RLR, Schweitzer ME, Recht MP, Wiggins GC, Regatte RR. MRI of the wrist at 7 tesla using an eight-channel array coil combined with parallel imaging: Preliminary results. Journal of Magnetic Resonance Imaging 2010;31(3):740-746.

[22].

Haase A, Frahm J, Matthaei D, Hanicke W, Merboldt K. FLASH imaging. Rapid NMR imaging using low flip-angle pulses. Journal of Magnetic Resonance 1986(67):258-66.

[23].

Koenig SH, Brown RD, Relaxometry of Tissue, in Encyclopedia of Magnetic Resonance. 1996, John Wiley & Sons, Ltd.

[24].

Koenig SH, Brown RD. Field-cycling relaxometry of protein solutions and tissue: Implications for MRI. Progress in Nuclear Magnetic Resonance Spectroscopy 1990;22(6):487-567.

[25].

Abragam A. The Principles of Nuclear Magnetism. 1961: Oxford University Press.

[26].

Momot KI, Takegoshi K, Kuchel PW, Larkin TJ. Inhomogeneous NMR Line Shape as a Probe of Microscopic Organization of Bicontinuous Cubic Phases. The Journal of Physical Chemistry B 2008;112(21):6636-6645.

118

Chapter 6: Higher field strength MRI scanning of long bones for generation of 3D models

[27].

Cavanagh J, Fairbrother WJ, Palmer AG, Rance M, Skelton NJ. Protein NMR Spectroscopy: Principles and Practice. 2nd ed. 2007, Burlington: Elsevier Academic Press.

119

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

Chapter 7 Step artefact caused by Magnetic Resonance Imaging of long bone 7.1 Introduction MR imaging of the musculoskeletal system is affected by various artefacts such as motion artefacts, chemical shift artefact, and magnetic susceptibility artefact (Some of these important artefacts have been discussed in Section 2.3.7). The motion artefacts (also referred to as the „movement artefact‟) occur due to the random or periodic movements of anatomical structures, resulting in blurred images and inaccuracies to the 3D models reconstructed from such image data. In an initial study, the supervisory team observed a step in the 3D model reconstructed from a data set obtained from the lower limb of a human volunteer that might have resulted from the volunteer moving the leg between two successive scanning stages [26]. In orthopaedic implant design, these artefacts can affect the design of anatomically well-fitting implants or their accurate validation.

Figure 7.1: The step artefact caused by volunteer moving the leg between two successive scanning stages

121

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

Motion artefacts due to periodic movements can be eliminated by synchronising the data acquisition with the movement, or by using post processing techniques. However, artefacts due to random movements cannot be eliminated with such techniques though radial K-space techniques are now available on clinical scanners to combat such motion artefacts to some degree. Since the step artefact has been observed in the reconstructed 3D models, the artefact can be eliminated using a 3D model aligning technique such as the iterative closest point (ICP) algorithm. This was successfully used in this study. This chapter is focused on correction of the step artefact of 3D models based on MRI. Section 7.2 will discuss the literature relevant to motion artefacts. Section 7.4 briefly introduces the methods used in the study and section 7.4 presents a summary of the study.

7.2 Motion artefact of MRI Motion artefacts are one of the challenges that researchers have faced when MRI is used for 3D reconstruction of long bones. This manifests as signal misregistration along the phase encoding direction, and the appearance may vary with the type and rate of the movement [40, 44]. The artefacts are caused by tissue excited at one location producing signals that are mapped to a different location during the data acquisition [40]. Motion artefacts in MR imaging are basically of two categories. The first category is the motion artefacts that occur due to periodic movements such as respiration, heart beat or flow of blood and cerebrospinal fluid. The second category is due to random movements such as the movements occur by the person‟s inability to keep the limbs still for long scanning duration or muscle contraction due to nerve stimulation from rapid change of the imaging gradients. The motion artefacts due to periodic movements have minimum or no effect for scanning of long bones of lower limbs, although scanning of the upper limbs might be affected by respiratory movements. The artefacts due to random movements, however, affect the MR imaging of the long bones of lower limbs. The lateral shift of the 3D models is one of the artefacts resulting from random movements of the lower limb. Due to the limitation of scanning length caused by the non-uniform magnetic field, the scanning of a long bone (e.g. femur or tibia) is conducted in several segments (Figure 7.2). 122

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

Figure 7.2: MRI scanning of human lower limb with five scanning segments to scan the complete limb Correction of the motion artefacts is generally achieved by synchronising the data acquisition with the movement, or by post processing the data; however, this is feasible only in the case of periodic movements [156, 157]. Artefacts due to random movements are hard to correct and different techniques have to be used, depending on the type of artefact. Immobilisation of the limb is one of the practices that can be used in the clinic; however, the muscle contractions due to the nerve excitations from RF waves cannot be prevented. Since this study is focused on correcting the lateral shift of 3D models, the use of 3D modelling technique is possible. The ICP algorithm is a robust method used to align 3D surfaces utilising the geometric features [158]. The ICP algorithm and the 3D -3D aligning process are described in Section 3.5.

7.3 Aims of the study This study aimed at correcting the step artefact that occurs due to the random movement of the lower limb, using the robust ICP algorithm based 3D modelling techniques.

7.4 Methods Five ovine hind limbs amputated from the pelvic and the ankle joint were used with intact soft tissue. The statistical sample size analyse shows that five samples would detect a difference of 0.07 mm with standard deviation of 0.02 for 80% power. The femora of the sheep hind limbs were scanned using a 3T MRI scanner with a customised protocol. Scanning was basically conducted to simulate the lateral shift artefact incurred by the random movements, which were achieved by shifting the bone laterally after scanning the first half of the bone. The artefact was corrected 123

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

using 3D modelling techniques to align the 3D models reconstructed from the two halves of the scanned bone. In addition, the errors resulting from the table movement were also quantified. A detailed description of the methods is available in the paper presented at the end of this chapter.

7.5 Results When the models with the corrected shift artefact were compared to the reference models, an average error of 0.32 ± 0.02 mm was generated. The 3D models reconstructed from the single MRI scan generated an error of 0.25 ± 0.02 mm. A detailed results section is available in the paper presented at the end of the chapter.

7.6 Summary, discussion and conclusion The motion artefacts occurring as a result of random movements play an important part when MRI data is acquired from long bones (mainly) for 3D reconstruction of long bones. Such an artefact causes the 3D models to have a step between two successive scanning segments. Unlike the movement artefacts due to periodic movements, the artefacts due to random movements cannot be eliminated by synchronising the data acquisition or by post processing techniques. Since the artefact is observed once the 3D model is reconstructed, a 3D surface aligning method is feasible to correct the artefact. The ICP algorithm is a robust and widely used method for 3D-3D alignment and was successfully implemented in this study for the correction of the step artefact. The results show that the geometric deviation of the corrected model is within the accepted accuracy levels for implant design. This error was slightly higher than the error obtained for the MRI based model reconstructed from the single scan. This residual error might have resulted from the slight mal-alignment between proximal and distal halves models. Statistical analysis of the sample size showed that this residual error (0.07 mm) with standard deviation of 0.02 could be detected statistically with the sample size of five. The study showed that by using the ICP algorithm, the step artefact observed in the 3D models of long bones can be corrected with sufficient accuracy to allow researchers to design orthopaedic implants using the 3D models generated from MRI. The present study utilised one simulated lateral shift; however, the human long bones have to be scanned in at least three stages, resulting in two lateral shift 124

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

artefacts. This might introduce higher errors to the corrected 3D models. Therefore, further validation of this method with human long bones has to be conducted before using it for correction of artefacts in human bone models. In this study the correction of the artefact was performed manually, using commercially available software and this is a labour intensive process. Hence, automatic processing to correct the artefact will be desirable in future.

125

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

7.7 Paper 5: Correction of step artefact associated with MRI scanning of long bones (Submitted – under review)

126

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

Correction of step artefact associated with MRI scanning of long bones 1

Kanchana Rathnayaka, 2Gary Cowin,

1,3

Michael A Schuetz, 4Tony Sahama, 1Beat

Schmutz 1

Institute of Health and Biomedical Innovation, Brisbane, QLD, Australia

2

University of Queensland, St Lucia, QLD, Australia

3

Princess Alexandra Hospital Brisbane, QLD, Australia

4

Queensland University of Technology Brisbane, QLD, Australia

Submitted to Journal: Medical Engineering and Physics Manuscript ID: MEP-D-11-00529 Corresponding Author: Dr. Beat Schmutz 60 Musk Avenue Kelvin Grove QLD 4059, Australia Email: [email protected]

127

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

Abstract Magnetic resonance imaging (MRI) has been shown to be a potential alternative to computed tomography (CT) for scanning of volunteers for 3D reconstruction of long bones, essentially avoiding the high radiation dose from CT. In MRI imaging of long bones, the artefacts due to random movements of the skeletal system create challenges for researchers as they generate inaccuracies in the 3D models generated by using data sets containing such artefacts. One of the defects that have been observed during an initial study is the lateral shift artefact occurring in the reconstructed 3D models. This artefact is believed to result from volunteers moving the leg during two successive scanning stages (The lower limb has to be scanned in at least five stages due to the limited scanning length of the scanner). As this artefact creates inaccuracies in the implants designed using these models, it needs to be corrected before the application of 3D models to implant design. Therefore, this study aimed to correct the lateral shift artefact using 3D modelling techniques. The femora of five ovine hind limbs were scanned with a 3T MRI scanner using a 3D VIBE based protocol. The scanning was conducted in two halves, while maintaining a good overlap between them. A lateral shift was generated by moving the limb several millimetres between two scanning stages. The 3D models were reconstructed using a multi threshold segmentation method. The correction of the artefact was achieved by aligning the two halves using the robust iterative closest point (ICP) algorithm, with the help of the overlapping region between the two. The models with the corrected artefact were compared with the reference model generated by CT scanning of the same sample. The results indicate that the correction of the artefact was achieved with an average deviation of 0.32 ± 0.02 mm between the corrected model and the reference model. In comparison, the model obtained from a single MRI scan generated an average error of 0.25 ± 0.02 mm when compared with the reference model. An average deviation of 0.34 ± 0.04 mm was seen when the models generated after the table was moved were compared to the reference models; thus, the movement of the table is also a contributing factor to the motion artefacts.

128

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

Introduction Magnetic resonance imaging (MRI) is theoretically designed to scan the soft tissues utilising the hydrogen nuclei as the source of signal. In recent studies, it has been shown to be a possible alternative to computed tomography (CT) for scanning of long bones [1, 2]. This alternative provides researchers designing orthopaedic implants with an opportunity to acquire long bone image data from the young healthy human population, who represent nearly half of all trauma patients, without having to expose them to the ionising radiation of CT [3]. However, MRI still suffers from some limitations such as very long scanning times, motion artefact and poor contrast in certain anatomical regions. Of these limitations, the motion artefact is crucial as it reduces the accuracy of the 3D models reconstructed from such image data [2]. A lateral shift has been observed in the 3D models reconstructed from data sets; this is believed to occur as a result of random patient movements [4]. The design of an orthopaedic implant needs accurate 3D representations of the relevant bone geometry. The current gold standard for acquisition of data for this purpose, CT, exposes a person to a high dose of ionising radiation. This exposure limits CT to the scanning of cadaver bones which are, in most cases, more than 60 year old. Since most of the patients who have been implanted with a plate or intramedullary nail are from the younger population, the implants need to be designed to suit this age group. For this purpose, there is an urgent need for the acquisition of data from this younger population. MRI is a versatile alternative for this purpose as there are no radiation hazards involved in MRI scanning. The poor contrast of certain anatomical regions in the MRI scanning of long bones can be overcome to some extent by using a higher field MRI scanner [2]. However, artefacts due to random movements of a subject remain a problem which needs to be addressed in order to utilise the models for the intended application. The motion artefacts occur when the protons of the tissue sample being scanned excited at one site are misregistered to another region of the image during the data acquisition [5]. This results in repeated reconstruction of the moving structures along the phase encoding direction [6]. The motion artefacts are of two types: the artefacts due to, periodic movements and the artefacts due to random movements [6]. The motion artefacts due to regular, periodic movements occurs (mainly) as a result of 129

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

respiratory movements, flow of blood in vessels or peristalsis. These artefacts mainly affect the MRI scanning of the relevant anatomical regions. Although respiratory movements might have an impact on the scanning of long bones of the upper limb, they have minimal or no effect on the scanning of the long bones of the lower limbs (e.g. the femur and tibia). The artefacts resulting from random movements may be due to nerve excitation during the scanning, or the patient randomly moving the limb. Due to the limited linearity of the gradients and B0 field of the MRI scanner, the lower limb of a subject has to be scanned in several stages (usually four to five). A preliminary investigation conducted by Schmutz et al. [4] using a MRI scanner has shown that the movement of the subject between these scanning stages produces a lateral displacement/shift in the final 3D model. The artefacts that result from periodic movements can be minimised by using various scanning protocols that synchronise the movement with the data acquisition or by using post processing/filtering techniques [7-10]. The artefacts due to random movements, on the other hand, cannot be eliminated easily by synchronising or post processing techniques of the image data. This can be achieved to some extent by immobilising the subject; however, immobilising a limb for ~60 minutes is not easily achievable. With regards to 3D model reconstruction, it can be achieved by 3D modelling techniques in which 3D models from the consecutive scanning stages can be aligned by using an iterative closest point (ICP) algorithm based technique [11]. The ICP algorithm is a widely used 3D-3D registration technique and has shown a high accuracy for translational as well as rotational alignments of 3D models. Lee et al. [1] conducted a preliminary registration test using the ICP algorithm in which a part of the bone model separated from the original model was matched perfectly to its original full model. In another study, the ICP algorithm was able to register a CT derived model to a real patient‟s model with an average error of 0.079 ± 0.068° for rotation and 0.12±0.09 mm on translations [12]. The ICP algorithm guarantees convergence to a local minimum from any given transformation of the data point set [11]. However, the obtained local minimum may not be the desired global minimum, as it depends on the initial registration. While the ICP algorithm has been used in numerous studies for aligning bone models, the effect of its initial position on the optimal global alignment has not yet been reported.

130

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

This study aimed to correct the lateral shift artefact that is associated with MRI scanning of long bones using the ICP algorithm for aligning the models. The dependency of the ICP algorithm on the initial position of the 3D surfaces to register them was also investigated.

Methods MRI scans of five ovine femora (Average age = 7 years and average weight = 49 kg) obtained by scanning five intact sheep hind limbs were used for the study. A 3T MRI scanner with the following imaging protocol (Table 1) and the body matrix coil was used. Table 1 MRI Protocol used for scanning of ovine femora Parameter Instrument Field Strength In plane resolution Slice thickness TE TR FA Image sequence Number of Averages

Value Siemens Trio tim 3T 0.47 mm × 0.47 mm 1 mm 1.83 ms 11 ms 10° 3D VIBE 1

Scanning of the femora was conducted using the setting described below. With the exception of the first step, the sample was scanned in two halves (proximal and distal) in all steps, while maintaining an approximately 7 cm overlap between the proximal and distal halves (Figure 1). 1. The femur was positioned in the centre of the magnet and a complete scan was obtained using a single field of view (FOV) (Figure 1a). 2. The femur was positioned in the centre of the magnet and the scanning was conducted in two halves using two FOVs without moving the table (Figure 1b). 3. The scanning was conducted in two halves using the same FOV used in the previous step; however, the table was moved such that the centre of the lower or upper half of the sample moved to the centre of the magnet.

131

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

4. The sample was scanned in two halves using the same FOVs as used previously and moving the table, as described in Stage 3. After the distal half was scanned, the proximal end of the specimen was shifted laterally to simulate the lateral shift caused by a volunteer moving their leg. Then, the proximal half was scanned (Figure 1c).

Figure 1: a - samples scanned with a single scanning segment; b – samples scanned in two segments without moving the table; c – samples scanned in two segments with a translated proximal segment (right) caused by the lateral shift of the specimen. 3D models of bones were reconstructed from all MRI data sets using the multithreshold segmentation method previously developed by the author [13]. This method combines a multilevel threshold approach with a method of selecting an appropriate threshold level for a particular anatomical region of a long bone. Two threshold levels were used for the two anatomical regions: the distal/proximal region and the diaphyseal region. Most of the articular regions were segmented manually

132

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

due to the presence of a number of different soft tissue types in the bone–soft tissue interface at those regions. The five femora were also scanned with a CT scanner as the reference standard against which to compare the MRI based models. A Toshiba 4 slice helical CT scanner was used with kVp = 120, mAs = 50, in plane resolution = 0.35 × 0.35 mm and slice spacing 0.5 mm. The CT data was segmented using the Canny edge detection method previously investigated by the author. The pair of 3D models reconstructed from the scans obtained without moving the table (Step 1) was used to quantify any displacement that might have occurred from the data acquisition process. The pair of 3D models reconstructed from the scans obtained after moving the table (Step 2) was used to quantify any displacement that might have resulted from movement of the table. The correction of the lateral shift that had been simulated during the scanning process was conducted using the ICP algorithm built into Rapidform 2006. The two 3D models reconstructed from scans of two halves of the bone were roughly aligned using the „Shell Trackball‟ tool in Rapidform 2006. The „Shell Trackball‟ tool allows translation of the model in any of the x, y and z directions and rotation around x, y and z axes. Only the distal half model was moved, while the proximal half model was kept locked in the 3D space of Rapidform 2006. After the rough alignment was carried out, the fine registration function that is based on the ICP algorithm was used for the final alignment of the models. For its operation, the ICP algorithm requires an overlapping region between the corresponding halves of the 3D models to be aligned (Figure 2 a & b). After alignment, the geometric deviation between two overlapping regions of the 3D models of two halves was measured using a point to point comparison method built into Rapidform 2006. Then the models of the two halves were merged using functions built into Rapidform 2006 to obtain the complete 3D model of the bone (Figure 2 a & b). The complete 3D models obtained without the table movement, with the table movement, and with corrected shift artefact were compared with the CT based reference model using the point to point comparison method built into Rapidform 2006 (Figure 2 c & d). Before the comparison, the 3D models were

133

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

aligned to the reference model using the fine registration (ICP based) function of Rapidform 2006.

Figure 2: a - 3D models of distal and proximal halves before the correction of the artefact, b – the artefact has been corrected by aligning the two halves, c - the corrected model (brown) is aligned with the reference model (blue), d - the colour map showing the differences between the corrected model and the reference model. The minimum overlap length that was required to accurately align the 3D models of the distal and proximal halves were determined prior to the correction of the artefact through the following procedure. A femur was MRI scanned two times using the same imaging protocol, shifting the sample from one end in the second scan to simulate a lateral shift. Two 3D models of the distal and proximal halves of the femur were reconstructed so that there was an approximately 7 cm overlap between the two models (Figure 3). This pair of models was copied 13 times. The models were then split so that each pair of models had varying overlap starting from 1cm, 134

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

and increasing in 0.5cm increment. In each pair of models, lateral shift was corrected using the same procedure used to align models based on the ICP algorithm, as described above. For each set, the corresponding two halves were then merged and compared against the reference model that was generated from a CT scan of the bone.

Figure 3: Overlapping region with reference planes created to divide the models. According to the results obtained (Figure 4), it can be deduced that the overlap of more than 4.5 cm produces acceptable alignment. Therefore, in this study, a 4.5 cm overlap was maintained between proximal and distal halves of all the reconstructed

0.10

0.08 0.06 0.04 0.02 0.00 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0

Average deviation (mm)

3D models.

Overlap (cm) Figure 4: Average deviations obtained for different overlapping regions of the 3D models. The dependency of the ICP algorithm on the initial positions of the 3D models for its alignment was investigated using two MRI based 3D models (proximal and distal halves) and their reference model. The proximal half of the MRI based 3D models 135

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

was placed in three positions (5mm, 10 mm and 15 mm apart from the distal half of the models) in each of the x, -x, y and -y directions (Figure 5), generating a total of twelve positions around the distal half for the MRI based models. The models were not positioned more than 15 mm apart, as the software‟s ICP based function was not able to align the models with more than a 15 mm distance between the two models. Then the function based on the ICP algorithm was used to align the proximal and distal halves of the MRI models. The two models were then merged and compared with the reference model for geometric deviations.

Figure 5: Proximal half of the MRI based model (blue) positioned in X and Y axes around the distal half of the MRI based model (red). Statistical differences between the average deviations of the models obtained from various scanning methods and the single scan model were calculated using one way ANOVA. The level of statistical significance was set to p ≤ 0.05. The validation was performed using PASW Statistics 18 software package.

Results When the geometric deviations between the overlapping regions of two halves were measured, the 3D models obtained without any table movements showed 0.18±0.11 mm average deviation. When it was measured in the models obtained after the table had moved, an average deviation of 0.49 ± 0.10 mm was obtained (Figures 6 & 7). After correcting the lateral shift artefact, the average deviation between the two overlapping regions was 0.05±0.01 mm.

136

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

0.7 0.6

Deviation ± SD (mm)

0.5 0.4 0.3 0.2 0.1 0 No table movement

With table movement

With corrected lateral shift artefact

Figure 6: Average deviations between overlapping regions of the models.

Figure 7: The lateral shift between the two 3D models obtained after the table was moved [A part of the distal model (pink) has been removed to show the displacement]. After merging the two halves, the obtained complete 3D models were compared to the reference models. The models obtained from two scans but without any table movements generated an average deviation of 0.26 ± 0.02 mm (Figure 8). The models obtained after the table had been moved presented an average deviation of 0.34 ± 0.04 mm (Figure 8), and the models with corrected lateral shift artefact 137

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

presented an average deviation of 0.32 ± 0.02 mm when compared to the reference models. The 3D models reconstructed from the MRI data that was obtained in a single scan showed an average deviation of 0.25 ± 0.02 mm when compared with the

Deviation ± SD (mm)

reference models. 0.50 0.45 0.40 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00

*

* *

*

No table movement

With table movement

With corrected lateral shift artefact

Single scan MRI

Figure 8: Average deviations between the complete models and the CT based reference standards (* = statistically significant). The results obtained for the investigation carried out to determine the dependency of the ICP algorithm on initial alignment of the models presented similar average deviations for the 12 positions (Table 2). Table 2 The accuracy of the ICP algorithm in aligning the 3D surfaces which have different initial alignments X

Axis Initial deviation (mm) Maximum (mm) Average (mm) SD Axis Initial deviation (mm) Maximum (mm) Average (mm) SD

138

5

10

-X 15

2.54831 0.31716 0.25116

2.55382 2.55255 0.31729 0.31725 0.25187 0.25164

5 2.54655 0.31708 0.25094

Y 10 15 2.54958 2.55511 0.31719 0.31733 0.25134 0.25201

5

10

15

2.54727 0.31713 0.25099

2.55275 2.51663 0.31722 0.31662 0.25172 0.24708

5 2.54944 0.31715 0.25135

-Y 10 15 2.54584 2.54726 0.31710 0.31714 0.25081 0.25101

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

Discussion This study investigated a method of correcting the lateral shift artefact that occurred as a result of the random movements of the subject between two successive scanning stages. This random movement is considered as one of the motion artefacts that occur in MRI imaging of long bones, in which the scanning is performed in a number of stages. The correction of the artefact is important before the models are used in various applications. In this study, a method of correcting this artefact was proposed and validated using the robust ICP algorithm to align the overlapping regions of two models with the simulated lateral shift artefact. It is known that the accuracy of the final optimal alignment performed by using the ICP algorithm is dependent on the initial position of the 3D surfaces. The investigation performed in this study, utilising two halves of a long bone, showed that the ICP algorithm based aligning method does not depend on the initial alignment of up to 15 mm for its registration process. The average errors obtained from this investigation were in the range of 0.31662 – 0.31733 mm with a standard deviation of 0.00018 mm between twelve measurements performed. Therefore, any effects on the alignment that might have been caused by the initial position of the models can be excluded. The minimum overlap required for the alignment of the two halves of the models can be as low as 4.5 cm, as determined by the investigation carried out in this study. The models obtained after the table was moved but without moving the specimen presented a higher error compared to the error obtained for the 3D models based on a single MRI scan. A lateral displacement of ~0.5 mm was visible in the 3D models reconstructed from two halves that were obtained after the table was moved. This lateral displacement is most likely to be caused by the mechanical instability of the moving table and/or by the slight movement of the sample resulting from the momentum contained in it. Generally, the scanning of a human long bone has to be performed by moving the table to cover the complete length of the bone and thus, any error generated due to the table movement is inevitable. The proposed method was able to correct the generated lateral shift artefact with an average error of 0.32 ± 0.02 mm between the model with corrected shift artefact and the reference model (CT based model). The error was within sub-voxel levels and 139

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

was slightly higher than the average error obtained for the models (scanned with two FOVs) reconstructed without moving the table (0.26 ± 0.02 mm) and the model obtained by using a single scan (0.25 ± 0.02 mm). The small residual error of the model with the corrected lateral shift artefact, compared to the model obtained with the single scan, is most likely to be the result of a slight mal-alignment between the proximal and distal halves. The average deviations between the model with the corrected shift artefact and the single scan model were significantly different statistically (p = 0.001); however, the difference between the model with the corrected shift artefact and the model obtained after moving the table was not statistically significant. Thus, using the proposed method, the lateral shift artefact can be corrected to an accuracy that is expected from clinical scanning where the table is moved. Generally the clinically acceptable tolerances for anatomically fitting fracture fixation plates are in the order of millimetres [14, 15]. Thus, the accuracy obtained, after correcting the shift artefact, is within the acceptable range for designing fracture fixation implants. The errors obtained for the single scan based MRI models (0.25 ± 0.02 mm) could have been the result of the manual segmentation that was performed in the articular regions of the MRI based models, and the larger slice spacing (1 mm) used in MR imaging compared to the 0.5 mm used in CT imaging. This error is consistent with the average error obtained for the comparison of MRI based models with the CT based models (0.23 mm) in a previous study conducted by the authors [2]. The articular regions are covered with a number of different types of soft tissue that exhibit different MRI properties. The contrast between those certain soft tissue types and the bone is generally not high enough for an accurate thresholding of the bone. Thus these regions were segmented manually, potentially introducing errors to those regions of the 3D models. The comparison of the models generated without table movements to the reference model presented an average error of 0.26 ± 0.02 mm. This error might have occurred mainly as a result of the segmentation process and the large slice spacing as mentioned in the previous paragraph [2]. However, the average deviation of 0.18 ± 0.11 mm that was measured between the overlapping regions suggests that there is a slight lateral deviation between two halves of the models that resulted from the scanning process. The exact reason for this deviation could not be determined. 140

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

The present study utilised only two scanning segments, resulting in one lateral shift artefact; however, when a human femur is being scanned, at least three segments have to be used; this results in two lateral shift artefacts. Thus, the error generated might be higher with a greater number of segments, compared to the present study. In addition, the correction of the artefact was performed manually and this is a labour intensive process. Therefore, automatic processing of the correction of artefact is desirable in future. The method proposed in this study was able to correct the lateral shift artefact of the 3D models based on MRI with acceptable accuracy for implant design. This was achieved using the robust ICP algorithm to align the 3D models using an overlapping region. The study also demonstrated that the ICP algorithm based function used in this study does not depend on the initial position of up to 15 mm for its alignment process. This allows medical engineering researchers to reconstruct accurate 3D models of long bones using MRI with minimum effect from the lateral shift artefact.

Acknowledgement This research was supported in part by Synthes GmbH. The last author has received an industrial scholarship from Synthes GmbH. The authors acknowledge the National Imaging Facility for providing 100% subsidised access to the 3T MRI scanner.

References [1].

Lee Y, Seon J, Shin V, Kim G-H, Jeon M. Anatomical evaluation of CT-MRI combined femoral model. Biomedical Engineering Online 2008;7(1):6.

[2].

Rathnayaka K, Momot KI, Noser H, Volp A, Schuetz M, Sahama T, Schmutz B. Quantification of the accuracy of MRI generated 3D models of long bones compared to CT generated 3D models. Medical Engineering & Physics 2011(in press DOI: 10.1016/j.medengphy.2011.07.027).

[3].

Henley G, Harrison JE, Serious injury due to land transport accidents, Australia 2006-07, in Injury research and statistics series. 2009, Australian Institute of Health and Welfare: quCanberra. p. 4-6.

141

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

[4].

Schmutz B, Volp A, Momot K, Pearcy M, Schuetz M. Using MRI for the imaging of long bones: First Experience. Journal of Biomechanics 2008;41(Supplement1):S188.

[5].

Brown MA, Semelka RC. MRI Basic principles and applications. 4th ed. 2010, New Jersey: John Wiley & Sons.

[6].

Peh WCG, Chan JHM. Artifacts in musculoskeletal magnetic resonance imaging: identification and correction. Skeletal Radiology 2001;30(4):179191.

[7].

Stadler A, Schima W, Ba-Ssalamah A, Kettenbach J, Eisenhuber E. Artifacts in body MR imaging: their appearance and how to eliminate them. European Radiology 2007;17(5):1242-1255.

[8].

Wood ML, Henkelman RM. MR image artifacts from periodic motion. Medical Physics 1985;12(2):143-151.

[9].

Cîndea N, Odille F, Bosser G, Felblinger J, Vuissoz P-A. Reconstruction from free-breathing cardiac MRI data using reproducing kernel Hilbert spaces. Magnetic Resonance in Medicine 2010;63(1):59-67.

[10].

Odille F, Cîndea N, Mandry D, Pasquier C, Vuissoz P-A, Felblinger J. Generalized MRI reconstruction including elastic physiological motion and coil sensitivity encoding. Magnetic Resonance in Medicine 2008;59(6):14011411.

[11].

Besl PJ, McKay ND. A Method for Registration of 3-D Shapes. IEEE Transaction on Pattern Analysis and Machine Intelligence 1992;14(2):239256.

[12].

Popescu F, Viceconti M, Grazi E, Cappello A. A new method to compare planned and achieved position of an orthopaedic implant. Computer Methods and Programs in Biomedicine 2003;71(2):117-127.

[13].

Rathnayaka K, Sahama T, Schuetz MA, Schmutz B. Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions. Medical Engineering & Physics 2010;33(2):226-233.

[14].

Schmutz B, Wullschleger ME, Noser H, Barry M, Meek J, Schuetz MA. Fit optimisation of a distal medial tibia plate. Computer Methods in Biomechanics & Biomedical Engineering 2011;14(4):359-364.

142

Chapter 7: Step artefact caused by Magnetic Resonance Imaging of long bone

[15].

Schmutz B, Wullschleger ME, Kim H, Noser H, Schuetz MA. Fit Assessment of Anatomic Plates for the Distal Medial Tibia. Journal of Orthopaedic Trauma 2008;22(4):258-263.

143

Chapter 8: Summary, conclusion and future directions

Chapter 8 Summary, conclusion and future directions 8.1 Summary and conclusion The overall objective of the research was to investigate the use of magnetic resonance imaging (MRI) to replace the current gold standard–computed tomography (CT)–so as to acquire long bone geometric data from healthy human volunteers. This data is required to design pre-contoured fracture fixation implants (plates and nails) to fit the anatomy of young patient age groups and patients from different ethnic groups. CT cannot be used for this purpose due to the involvement of high amounts of ionising radiation. With this overall objective, the study specifically aimed to: develop a simple and accurate segmentation method for segmentation of MRI and CT data of long bones; formally validate the geometric accuracy of the MRI and CT based 3D models of long bones with an appropriate reference standard; use higher field 3T MRI to improve the poor contrast of certain anatomical regions (which is a limitation of current 1.5T MRI scanners); and correct the step artefact in the 3D models caused by the movement of volunteers during the MRI scan. The reconstruction of 3D models of bones with accurate representation of the surface geometry requires using an accurate segmentation method. Currently available sophisticated segmentation methods are capable of segmenting relatively short bones with minimum user intervention; however, the accessibility of these methods by the general research community is limited due to the complex mathematics and programming involved. This study proposed and validated two relatively simple but accurate segmentation methods: multi-threshold segmentation and Canny edge detector based segmentation, which can be used to accurately segment the CT and MRI images of long bones. The former uses the popular intensity thresholding with 145

Chapter 8: Summary, conclusion and future directions

multiple threshold levels for regions of the bone that have different intensity levels. The threshold levels were calculated using the developed threshold selection method to minimise user dependent errors of selecting a threshold level. The latter uses the Canny edge detector which is already built into common image processing platforms (e. g. Matlab and IDL). Both segmentation methods were capable of segmenting outer and inner surfaces of ovine femora from CT images with high accuracy when compared with the reference standards. MRI, as an ionising radiation free imaging method, has shown potential for scanning of bones for reconstructing 3D models. This was formally validated for reconstructing 3D models of long bones with accurate surface geometry, using 1.5T MRI and CT scans of ovine femora. The state of the art dense triangular meshed surfaces generated from a contact mechanical scanner were used as the reference standard. Image segmentation of both CT and MRI data was conducted using the multi-threshold segmentation method developed in this study. Results showed that there was no statistically significant difference between the obtained MRI based 3D models and the CT based models. Compared to the diaphyseal regions, the articular regions of the MRI based 3D models presented lower accuracy. This is due to the poor contrast in those regions resulting from a number of different types of soft tissue with different MRI properties that surround the bone. Segmentation of MRI images takes longer than segmentation of the CT images, especially in articular regions; this is also labour intensive compared to CT images. In addition, MRI‟s very long scanning times make the images vulnerable to the artefacts caused by random movements of the subject. These factors might limit the use of MRI for reconstruction of 3D models of long bones. There are some promising approaches to addressing these current limitations of using MRI for scanning of long bones. Higher field strength MRI scanners promisingly offer higher signal to noise ratio (SNR) levels that can be used either to reduce the scanning time or improve the poor contrast in articular regions. Since the commonly used higher field strength MRI scanner in the clinical setting is 3T, in the present study, a comparison between 1.5T and 3T was conducted to quantify the improved image quality at 3T. The comparison using signal to noise ratio (SNR) of soft tissues and bone marrow, and contrast to noise ratio (CNR) of bone–muscle and bone–bone 146

Chapter 8: Summary, conclusion and future directions

marrow interfaces resulted in comparatively higher SNR and CNR levels for most of the regions of the femur and tibia. The increased contrast at 3T might improve the segmentation accuracy of the articular regions; however, according to the author‟s experience of segmentation, this only marginally reduces the segmentation time in comparison to the 1.5T images. Whilst SNR and CNR are increased at 3T, some of the artefacts may also be exaggerated. The magnetic susceptibility becomes more apparent at 3T and the chemical shift artefact is doubled due to the increased difference of the resonance frequency between water and fat molecules. Since the strength of the magnets is being increased over time, scanners with higher magnetic field (e.g. 7T) than 3T will potentially increase the image quality. However, increased SAR levels at higher magnetic fields will potentially limit their use for human imaging. The investigation showed that the longitudinal relaxation time (T 1) of the muscle was highly field dependent, while T1 of bone marrow was weakly field dependent. In both muscle and bone marrow, T1 increased at 3T. In contrast, the transverse relaxation time (T2) of muscle was not field dependent; however, the literature reports that T2 takes slightly lower values at higher magnetic field strengths. Increased T1 at 3T requires relatively higher TR values to be used to get the maximum intensity levels and this, in turn, increases the scanning time. In general, this can be compensated for by using fewer averages when data is acquired; however, in the present study, it is not possible as the number of averages used is one. In MRI imaging of long bones of lower limbs, the artefacts due to the random movements of the subject are relatively more prominent and important than those due to the periodic movements such as respiration and blood flow. As observed by the supervisory team, the random movements between two successive scanning stages causes a step in the 3D models reconstructed from such data sets of lower limbs. The artefacts due to the random movements cannot be eliminated by post processing or by synchronising the data acquisition. Immobilisation of the limb for about 60 minutes is also not achievable unless an invasive anaesthetic method is used. However, as the lateral shift artefact appears in the reconstructed 3D models, the robust ICP algorithm that has been widely used for 3D-3D registration of 147

Chapter 8: Summary, conclusion and future directions

surfaces was implemented for correcting this artefact when simulated in ovine femora. The resulting 3D models had sub voxel-level accuracy (voxel size = 0.35 mm2) when the surface geometry was compared to the reference standard. The study also indicated that the movement of the table makes a displacement in the data sets; however, this may not be important in clinical applications. Nevertheless, to minimise the geometric errors, the data acquisition of long bones for 3D reconstruction should consider this displacement caused by the table movement. In conclusion, magnetic resonance imaging, together with simple multi-level thresholding segmentation, is able to produce 3D models of long bones with accurate geometric representations. It is, thereby, a potential alternative scanning method where the current gold standard CT imaging cannot be used. However, there are a number of limitations such as long scanning times, long segmentation time, and movement artefacts that have to be resolved before employing MRI for this purpose.

8.2 Future directions This study successfully validated the accuracy of MRI to reconstruct 3D models from long bones using simple but accurate segmentation methods. The usability of 3T MRI scanners was also investigated, while 3D modelling techniques were used to correct the shift artefacts. However, there are a number of limitations or challenges that should be addressed in the future, before using MRI as an alternative to CT for imaging of long bones for 3D reconstruction. The segmentation methods described in this research may also be used in fields other than 3D reconstruction of long bones. Cardiac MRI and CT image segmentation is one such area where accurate segmentation is required for volumetric measurements. MR only radiotherapy planning is another aspect in the clinic that requires accurate segmentation of bone and soft tissue and these methods can potentially be employed for these purposes. With regards to image segmentation, segmentation of MRI images takes a considerably longer time compared to the segmentation of CT images. Even though 3T scanners are able to improve the contrast levels in articular regions, according to the author‟s experience, this only marginally reduces the segmentation time. Future studies might focus on automating the segmentation process to reduce the 148

Chapter 8: Summary, conclusion and future directions

segmentation time. In addition, the use of magnetic fields stronger than 3T may also improve the contrast levels, thus allowing faster segmentation, especially of the articular regions. In addition to the use of higher field scanners to improve the image quality, this may also be achieved with specially designed RF coils or imaging protocols. In the present study, the peripheral angiography (PA matrix) coil was used for imaging of the lower limbs; however; there are no RF coils currently available for scanning of upper limbs. Therefore, designing RF coils especially for scanning of long bones of the upper and lower limbs in a future study will improve the quality of MRI images of bones and, hence, the segmentation accuracy and time. Using imaging protocols such as the protocols designed for fat and water only imaging or protocols with ultra short TE (UTE) will potentially improve the CNR between bone and soft tissues. Furthermore, currently available imaging sequences are also mainly designed to scan soft tissues. Collaboration with manufacturers to design protocols for scanning of bones could also have an influence on improving image contrast of MRI of bone. The present study validated the correction of step artefacts in MR imaging of long bones using ovine femora which is relatively smaller than human femora. Therefore, this method has still to be validated using human long bones before using it to successfully generate 3D models of human long bones. Future studies can be conducted using fresh human cadaver bones and CT as the reference standard. According to the studies conducted using 1.5T and 3T MRI scanners, MRI of human long bones results in very long scanning times. This can potentially be shortened in the future by using higher magnetic fields (e.g. 7T). In addition, optimising the scanning protocol for different regions of the bone (e.g. use of larger slice spacing and low resolutions for diaphyseal region where the geometry is relatively simple) may also reduce the scanning time. Even though artefacts due to periodic movements are not prominent in the MRI scanning of long bones of lower limbs, scanning of upper limbs are affected by respiratory movements. Therefore, minimising periodic motion artefacts is also important in the long bone MRI of upper limbs. In this study, the lateral shift artefact was corrected using 3D modelling techniques by manually positioning the bone models. However, in future, automatic processing is desirable in order to reduce the 149

Chapter 8: Summary, conclusion and future directions

time taken for manual processing. In addition to motion artefacts, minimising the artefacts produced by magnetic susceptibility and the chemical shift may also be important in MRI scanning of bones, especially when high magnetic fields are used as these exaggerate the artefacts. Minimising or elimination of these artefacts is important for improving the accuracy of the implants designed using the MRI based 3D models.

150

Appendix 1

Appendix 1

Ethical approval for the study in Chapter 6

•

•

-

Queensland

,

Government

~

Royttl Brishant: o.mU \Voml!n·s Hospital Mctm Nonh Health Service District

Office of the Human Research Ethics Committees

Queensland Health Em1uioics 10:

Odcnc Pctco,;cn Coordinator

Phone: Fax: Our Rcf: E-mail

07 3636 5490 07 3636 5849 HREC/1 0/QRBW/1 41 IW\1'11-I'oluc''" healt h qld go,·.au

Or Kanchana Rathnayaka Queensland University of Technology Institute of Health & Biomedical Innovation 60 Musk Avenue Q 4059 Kelvin Grove

Dear Dr Rathnayaka, Re:

Ref N!l: HREC/10/QRBW/141: Comparative study of 3T MRI vs I.ST for the acquisition of 3D morphological bone data of the lower extremity

Thank you for submitting the above project for ethical and scientific review. This project was considered at the Royal Brisbane & Women's Hospital Human Research Ethics Committee (HREC) meeting held on I 9 April, 2010. I am pleased to advise that the Human Research Ethics Committee has granted approval of this research project on 13 May, 2010. HREC approval is valid for three (3) years from the date of this letter. This HREC is constituted and operates in accordance with the National Health and Medical Research Council 's (NHMRC) National Statement 011 Ethical Co11duct in Human Research (2007). NHMRC and Universities Australia Australian Code.for tlze Responsible Conduct of Research (2007) and the CPMPIICH Note.for Guida11ce on Good Clinical Practice. Attached is the HREC Composition with specialty and affiliation with the Hospital (A twchmc!Tt lj. You are reminded that this letter constitutes ethical approval only. You must not commence this research proJect at a site until separate authorisation from the District CEO or Delegate oftlwt site ftas been obtained. A copy t?fthis approval will also be sent to the Di.vtrict Research Governance Office (RGO). Please ensure you submit a completed Site Specific A!!·sessmeut (SSA) Form to the RGO for authorisation from tfze CEO or Delegate to conduct this research at the Royal Brisbane & Women's Ho.\pital ML'fm North District.

The documents reviewed and approved include: !!to'

/lo_•·t~f llri.