3D Reconstruction of Long Bones Utilising Magnetic ... · Abstract IV usability of MRI as a...

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

III

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

Abstract

IV

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

Abstract

V

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.

Keywords

VI

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

Contents

VII

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

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

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

2.3 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

Contents

VIII

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

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

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

2.4 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

3.8.1 Samples ............................................................................................. 32

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

Contents

IX

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

3.8.5.1 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

3.9 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

Contents

X

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

Contents

XI

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

List of figures

XIII

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

List of figures

XIV

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

List of tables

XV

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

Publications, presentations and awards

XVI


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


XVII

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, 21-

24 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.

Authorship

XIX

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

Acknowledgement

XXI

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.

Abbreviations

XXII

Abbreviations 13

C Carbon 15

N Nitrogen 1H Hydrogen

31P Phosphorus

3D Three dimensional

3T Three tesla

B0 Main magnetic field

BW Bandwidth

CAS Computer assisted surgery

CNR Contrast to noise ratio

CT Computed tomography

FA Flip angle

FOV Field of view

H2O Water

HU Hounsfield units

ICP Iterative closest point

M0 Net magnetisation vector

MHz Mega Hertz

MR Magnetic resonance

MRI Magnetic resonance imaging

Mt Transverse component of net magnetisation vector

Mz Longitudinal component of net magnetisation vector

NAV Number of signal averages

NMR Nuclear magnetic resonance

NPA Number of acquired partitions

NPE Number of acquired phase encodes

PMMA Poly-methyl methacrylate (Dental acrylic)

RF Radiofrequency

ROI Region of interest

SD Standard deviation

SNR Signal to noise ratio

SNRGER Signal to noise ratio for gradient echo sequence

SNRSE Signal to noise ratio for spin echo sequence

T1 Longitudinal relaxation time

T2 Transverse relaxation time

TE Echo time

TMS Tetramethylsilane

TR Repetition time

V Voxel volume

Chapter 1: Introduction

1

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].


2

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


3

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.


4

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.


5

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

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

7

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 x-

ray 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.


8

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).


9

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


10

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,

13C,

15N,

31P. 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).


11

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


12

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):

2

00

Bv 2.1

Where 0v 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


13

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 T1 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):

)( 0 ii rGBv 2.2

Where iv is the frequency of the spin at position ir , 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 frequency-

encoding) 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.


14

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 nickel-

titanium 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


15

relationship between SNR and other parameters of a MRI system for a spin echo

sequence and gradient echo sequence [43]:

Noise

SignalSNR 2.3

21 )1(0

TTETTRAVPAPESE ee

BW

NNNVBSNR 2.4

2

1

1

cos)1(

)1(sin0

TTE

TTR

TTR

AVPAPEGER e

e

e

BW

NNNVBSNR 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.


16

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


17

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:

00 BM 2.6

Where M0 = net magnetisation vector, = 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.


18

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.


19

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


20

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.

Chapter 3: Image processing and surface reconstruction

21

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


22

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


23

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.


24

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


25

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).


26

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.


27

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.


28

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.


29

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


30

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


31

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


32

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 under-

estimating the regions with different threshold levels when a single threshold level is


33

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).


34

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.


35

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 Value

Sensor 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.)

Scanning method Contacting, mesh-point height-sensing

Scanning pitch

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.)

Scanning speed 4-15 mm/sec. (1/8-9/16 in./sec.)

Exportable file formats DXF, VRML, STL, 3DMF, IGES, Greyscale, Point

Group and BMP

XY table size 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.


36

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


37

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.


38

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


39

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


40

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 Value

Resolution 0.03 mm × 0.03 mm

Slice spacing 0.03 mm

kVp 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


41

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.


42

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


43

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


44

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


45

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.


46

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.


47

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.


48

3.11 Paper 1: Effect of CT image segmentation methods on the

accuracy of long bone 3D reconstructions (published)


49

226-233

Contents lists available at ScienceDirect

Medical Engineering & Physics

ELSEVJER journal homepage: www.elsevier.com/locate/ medengphy

Effects of CT image segmentation methods on the accuracy of long bone 3D reconstructions

Kanchana Rathnayaka a, Tony Sahamaa, Michael A. Schuetza.b, Beat Schmutza .•

• Institute of Health and Biomedical Innovation. Queensland University of Technology, Brisbane, Australia b Department of Orthopaedics. Princess Alexandra Hospital. Brisbane. Australia

ARTICLE INF O ABSTRACT

Article history: 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

An accurate and accessible image segmentation method is in high demand for generating 3D bone models from CT scan data, as such models are required in many areas of medical research. Even though numerous sophisticated segmentation methods have been published over the years, most of them 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 implementations 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 ith 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 ith 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.

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 which the 3D models of long bones are generated. Two main steps are involved in generating 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 the 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 segmentation 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: fax: +61 7 3138 6030.

£-mail addresses: [email protected], [email protected] (B. Schmutz).

© 2010 IPEM. Published by Elsevier Ltd. All rights reserved.

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 the 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 automated techniques ( 17-24].

Manual segmentation/tracing 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 consuming 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 threshold 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 appropriate threshold level, various

1350-4533/$ - see front matter © 2010 IPEM. Published by Elsevier Ltd. All rights reserved. doi:1 0.1 016/j.medengphy.201 0.10.002

halla

Due to copyright restrictions, this article is not available here. Please consult the hardcopy thesis available from QUT Library or view the published version online at: http://dx.doi.org/10.1016/j.medengphy.2010.10.002

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

57

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.


58

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


59

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.


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.


60


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 non-

invasive. 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.


61

4.7 Paper 2: Quantitative fit assessment of tibial nail designs using

3D computer modelling (published)


62

Injury, lnt. j . Care Injured 41 (2010) 216-219

Contents lists available at ScienceDirect

Injury

ELSEVIER journal h omepage: www.elsevier.com/ lo catelin ju ry

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

Introduction

ABST R ACT

Intramedullary nailing is the standard fixation method for displaced diaphyseal fractures of the tibia in adults. The bends in modern tibial nails allow for an easier insertion, enhance the 'bone-nail construct' stability, and reduce axial malalignments of the main fragments. Anecdotal clinical evidence indicates that current nail designs do not fit optimally for 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 tibiae by utilising 3D computer modelling.

We used 3D models of two diffe rent tibial nail designs (ETN (Expert Tibia Nail) and ErN-ProximalBend, Synthes), and 20 er-based 3D cortex models of Japanese 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 mm2

) compared to the ETN (average 1044 mm2) . 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.

Intramedullary nailing is the standard fixation method for displaced 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/Weste 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 designs used in th is study. The results of a clinical study13 from one of the

hospitals involved in the multi-centre clinical tria l o f this nail confirmed the improvements5 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 Asian patients.

One important aspect of designing a new or improved implant shape 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 visual assessment. The force required for inserting the nail into the bone is

• 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:1 0.1 016fj.injury.2009.1 0.012

halla

Due to copyright restrictions, this article is not available here. Please consult the hardcopy thesis available from QUT Library or view the published version online at: http://dx.doi.org/10.1016/j.injury.2009.10.012

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

67

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.


68

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).


69

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 T1 weighted


70

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


71

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.


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


72

= 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.


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


73

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.


74

5.7 Paper 3: Quantification of the accuracy of MRI generated 3D

models of long bones compared to CT generated 3D models (in

press)


75

Medical Engineering & Physics xxx (2011) xxx-xxx

Contents lists available at Scie nceDirect

Medical Engineering & Physics

ELSEVIER journal homepage: www . elsev ie r .co m /locat e/medengp h y

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. Schuetza,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. 1 99 1pswich Road. Woolloongabba. Brisbane. QLD 4102, Australia

ARTICLE INFO ABSTRACT

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 fixation implants are increasingly being designed using accurate 3D models of long bones based on computer tomography (Cf). Unlike cr. magnetic resonance imaging (MRI) does 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 models 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.

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).

© 2011 IPEM. Published by Elsevier Ltd. All rights reserved.

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

halla

Due to copyright restrictions, this article is not available here. Please consult the hardcopy thesis available from QUT Library or view the published version online at: http://dx.doi.org/10.1016/j.medengphy.2011.07.027

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

83

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 (T1) 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


84

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:

2

0BS 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.

00 BNvN 6.2

22

4

N

SSNR 6.3

Where N = Noise level, 0v = 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].


85

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<0.01).

Gold et al. [147] conducted a study to calculate the spin relaxation times of 3T and

1.5T of the musculoskeletal system using five human volunteers. The T1 was

increased in the 3T MRI compared to the 1.5T MRI, while T2 was slightly decreased.

In the same study, SNR and CNR of muscles were compared using one volunteer

whose knee was scanned in both 1.5T and 3T. A sagittal, proton density weighted

fast spin-echo sequence was used with a TR and TE of 4000 ms and 14 ms

respectively. A coronal T1-weighted spin-echo sequence was used with TR and TE of

800 ms and 14 ms respectively. The parameters were identical at both field strengths.

SNR and CNR were calculated for cartilage, muscle, fat and synovial fluid. SNR of

muscle and fat were more than twice at 3T compared with 1.5T. CNR between

synovial fluid and cartilage at both long TR (4000 ms) and short TR (800 ms) had

increased at 3T compared with 1.5T. All the SNR measurements were higher at 3T

and the values were statistically significant. A pictorial comparison conducted by

Gold et al. has also shown that 3T MRI gives significantly better results for

musculoskeletal tissues [128].

In addition to the studies mentioned above, Lambert et al. showed that 3T MRI can

be used to detect rotator cuff tears; this was a follow up qualitative study and no

comparison was performed with 1.5T MRI images [138]. No studies that compare

1.5T and 3T bone–muscle interface in relation to generation of 3D models have been

reported.

6.3.1 Spin relaxation times and flip angle

Longitudinal relaxation time (T1) is highly dependent on the strength of the main

magnetic field and increases with the increase of the main magnetic field [23, 147].

The reported increase of T1 relaxation times at 3T are: 20 - 22% for fat [131, 147],

20% for skeletal muscle, 62% for gray matter, 42% for white matter in the brain and


86

41% for liver [147, 148]. Previous measurements of relaxation times have shown 70-

90% increase of T1 at 4T than at 1.5T [139]. This increased longitudinal relaxation

time at 3T requires the TR to be increased to obtain the same imaging contrast as at

1.5T, especially for T2 and proton density weighted images; this, in turn, increases

the imaging time at 3T. This also decreases the SAR values and allows more slices

per scan.

Transverse relaxation time (T2) is relatively less dependent on the main magnetic

field. T2 slightly decreases with the increase of the main magnetic field; thus, similar

contrast levels may be obtained at slightly shorter TEs at 3T compared to 1.5T [141].

Gold et al. have reported that the decrease of T2 is about 10% for muscle and 19%

for fat (marrow and subcutaneous) [147].

For a given TR at a certain T1, the optimum flip angle should be used to obtain the

maximum signal. This is called the „Ernst angle‟, and is given by the following

equation:

1cosTTR

e 6.4

Where is the flip angle, TR = repetition time, and T1 = longitudinal relaxation time

of the tissue.

6.3.2 Fat suppression

Fat suppression is a technique used in MRI to suppress the signal from normal

adipose tissue to reduce the artefacts or to characterise the tissue [149]. This is

usually achieved by sending pulse to suppress the resonance frequency of the fat

tissue. Since the difference of the resonance frequencies of water and fat is 220Hz at

1.5T, a pulse used to suppress the signal from fat will partially or completely

suppress the signal from water molecules. As this difference becomes 440Hz at 3T,

fat suppression can be achieved easily without suppressing the signal from water

molecules [141].


87

6.3.3 Magnetic susceptibility at 3T MRI

Magnetic susceptibility occurs due to the change of the static magnetic field in the

presence of materials such as bones, metals, certain blood products and air. Due to

this variability of the static magnetic field strength, spins precess at different

frequencies. High magnetic fields are highly sensitive and prone to magnetic

susceptibility more than 1.5T fields [150]. Magnetic susceptibility causes a signal

loss, leading to geometric distortions of the region. This will affect the localisation of

certain anatomical structures, not only in clinical practice but also in the 3D models

generated using such a data set.

The degree of magnetic susceptibility can be minimised with various practices.

Increasing bandwidth decreases susceptibility artefact; however, this is at the

expense of SNR (Equations 2.4 & 2.5 P15). By doubling the bandwidth, SNR is

decreased by about 40% [23]. This can also be achieved by decreasing echo time

(TE). Decreasing the voxel volume is another method which can be used to reduce

the susceptibility artefact. This is usually achieved by using higher special resolution

with thinner slices.

6.3.4 Chemical shift at 3T

Chemical shift is the difference in precessional frequency conferred by 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, or

TMS). The chemical shift is present in any tissue, but the electron cloud in the fat has

a major effect as fat is an abundant homogenous tissue in the body [141]. Due to this

frequency difference between lipid/fat and water, protons from the same tissue

location map to different positions in the reconstructed image.

The difference of the precessional frequencies between water and fat is

approximately 220Hz at 1.5T. As this frequency is directly proportional to the main

magnetic field, it becomes approximately 440 Hz at 3T [128, 141, 151]. This doubles

the chemical shift between water and fat, doubling the number of misregistered

pixels and worsening the artefact. The chemical shift can be minimised by increasing

the receiver bandwidth of the scanner; for example, doubling the BW will decrease

the chemical shift to the same as at 1.5T. However, doubling the receiver BW

decreases the SNR by a factor of square root of 2.


88

6.3.5 MRI safety at 3T

As discussed in Chapter 2, MRI does not have known direct hazards from the main

magnetic field. However, the RF energy absorbed is converted to heat inside the

body. To prevent the excessive heat generation inside the patient, specific absorption

rates (SAR) are monitored and maintained within certain limits. For low field

strength scanners, SAR levels do not offer any limitations; however, high field

strength scanning has certain effects from increased SAR levels. SAR is proportional

to the square of the strength of the main magnetic field, as given below (6.5).

Therefore, at 3T, the SAR level is four times that at 1.5T:

DBSAR 2

0 )( 6.5

Where SAR = Specific absorption rate, B0 = strength of the main magnetic field, =

flip angle and D = the duty cycle.


The study aimed to quantitatively compare the quality of the images obtained at 1.5T

to those acquired at 3T. The study also quantified the change of spin relaxation times

at 3T compared to 1.5T MRI.

6.5 Methods

6.5.1 Samples

Right legs of five human volunteers were scanned in 1.5T and 3T MRI scanners. The

calculation of sample size showed that for 80% of power a total of 20 samples need

to be used to detect a difference of 1.76 with standard deviation of 1.00. However,

this study had a sample size of five due to the very long processing time. Using the

sample number of five the difference that can be detected is 3.55 with the same

standard deviation.

6.5.2 Measuring the quality of MR images

Generally, the quality of an image is assessed using signal to noise ratio (SNR).

Signal to noise ratio is the proportion between the signal and the noise of the image;

thus, the higher the SNR, the better the image quality. SNR is one of the important


89

measures of the performance of a MRI system, as discussed in Section 6.2 [152]. The

equation below (Equation 6.6) gives the signal to noise ratio of any signal:

Noise

SignalSNR 6.6

In images, average pixel intensity of a sample from the ROI is taken as the signal and

the noise is calculated from the background of the image. Noise can either be the

average of the background intensity or the distribution of signal (Standard deviation

of the background intensity). In both cases, equal results can be obtained when

proper conversion factors from noise statistics are used [153]. With regards to SNR

of MRI, this method can be applied to calculate the SNR from a single MR image.

However, methods which utilise two or a series of images are also available [152,

154].

When SNR for a single MR image is calculated, a statistics derived factor

)4/(2 is introduced to the background noise [155]; thus, the equation can be

modified as in Equation 6.7:

noise

tissue

STD

MSNR

4

2

6.7

Where SNR = Signal to noise ratio, Mtissue = Average intensity of desired tissue type,

and STDnoise = Standard deviation of background noise.

Contrast to noise ratio (CNR) can be used to assess the contrast (intensity difference)

between two ROIs of image. This is an important measure used to assess or compare

the quality of MR images where higher contrast is required to identify certain tissue

types from the others; as an example, to identify the cortical bone from the

surrounding muscle tissue, the contrast between bone and muscle should be

considerably higher. In the basic form, CNR is the ratio between intensity difference

of two signals and the noise level (Equation 6.8):

Noise

SignalSignalCNR 21 6.8


90

As per SNR, the noise level can be calculated using the SD of the background; thus,

with addition of the noise statistics derived conversion factor, the CNR can be

calculated as in Equation 6.9:

noise

tissuetissue

STD

MMCNR

4

2

21 6.9

Where Mtissue1 = Average intensity of tissue type-1, Mtissue2 = average intensity of

tissue type-2, and STDnoise = Standard deviation of the average intensity of muscle.

6.5.3 Quantification of spin relaxation times

As per the discussion earlier in this chapter, T1 is highly field-dependent and, thus,

there is a need to use higher TR values in order to obtain comparable SNR values as

at 1.5T. Higher TR values, however, result in increased imaging time and, therefore,

there is a need to optimise the imaging protocol of 3T to obtain a better contrast

without losing the SNR. In order to calculate the T1 and T2 values of the muscle and

bone marrow, a series of images of the centre of the femur was obtained with varying

TR and TE values (Table 6.1) using 3T and 1.5T MRI scanners, while keeping the

other parameters constant.

Table 6.1 TR and TE values used for the MRI scanning at 1.5T and 3T for

calculation of T1 and T2

TR (ms) TE (ms)

1.5T 10 12 15 20 30 50 1.5T 4 5 6 8 10 12

3T 10 14 20 30 50 100 3T 4 5 7 9 12 14.8

The effect of different FAs on SNR was investigated by scanning the mid femoral

region of a volunteer with varying FAs (Table 6.2), while keeping all the other

parameters constant.

Table 6.2 Different flip angles used for scanning

Flip angle

1.5T 7 9 12 15

3T 7 9 12 15


91

T1 and T2 spin relaxation times were calculated for muscle and bone marrow

compartments from the images obtained, as described above. T1 and T2 were

determined, as indicated by the equations below:

s

S

NRTR

RF

NRTR

e

eSTRS

1

1

cos1

1)( 0 6.10

2

0)(RTE

eSTES 6.11

Where NS = 20 was the number of slices and R1 and S0 were the adjustable fit

parameters. The fitted value of R1 was taken as the longitudinal relaxation rate, 1/T1,

in the respective voxel. The fitted value of R2 was taken as the transverse relaxation

rate, 1/T2, in the respective voxel. SNR was also calculated for muscle, cortical bone

and bone marrow, while CNR was calculated at the muscle–cortical bone and bone

marrow–cortical bone interfaces. The results are included in the paper presented at

the end of this chapter.

6.5.4 Comparison of 1.5T and 3T imaging of musculoskeletal system

This section of the chapter describes the methods used for quantitative comparison of

the image quality of MRI at 1.5T to 3T, with special consideration to generation of

3D models of long bones. Comparison of the image quality was carried out using the

measures described in the previous section of this chapter dealing with SNR and

CNR.

The right legs of five healthy human volunteers were scanned using a 1.5T and 3T

MRI scanner from the same manufacturer. Identical protocols (Table 6.3) were used

for the scanning. Two RF coils (PA and Body matrix) were used to cover the lower

limbs completely (Figure 6.1).


92

Table 6.3 The protocols used for MRI scanning

Parameter 1.5T 3T

In plane resolution 0.45 mm ×0.45 mm 0.45 mm ×0.45 mm Slice thickness 1 mm 1 mm

TR 11 ms 11 ms

TE 4.66 ms 4.66 ms

Flip angle 7° 7° Number of averages 1 1

FOV 176×256 176×256

Matrix size 352×512 352×512 Image sequence 3D FLASH 3D FLASH

Manufacturer Siemens Siemens

Model Magnetum Avanto Trio Tim

RF Coils PA & Body Matrix PA & Body Matrix

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

In order to cover the lower limb completely, five scanning stages (Figure 6.2), each

containing 256 slices, were required, depending on the height of the subject. These

were positioned so that there were 66 slices of overlap between two successive

scanning stages. These overlapping regions are later used for the alignment of the 3D

models.


93

Figure 6.2: Positioning of the field of view (FOV) on volunteer‟s leg

SNR and CNR were calculated for the proximal, diaphyseal and distal regions of

both tibia and femur. Image slices were obtained from nearly identical locations of

both 1.5T and 3T data sets. Five samples from each of muscle, cortical bone and

bone marrow were obtained using a customised Matlab script. The calculated SNR

and CNR values were used to compare 1.5T and 3T MR images. A detailed methods

section is available in the paper presented at the end of this chapter.

6.6 Results

The comparison between 1.5T and 3T images of the femora produced the following

results. In the mid diaphyseal region, CNR and SNR of muscles were higher for 3T

compared to 1.5T. In the proximal diaphyseal region, CNR and SNR of muscle at 3T

were slightly higher than 1.5T. In the distal diaphyseal region, CNR and SNR of the

other soft tissues were slightly higher at 3T than 1.5T; however; CNR and SNR of

muscles were slightly higher at 1.5T compared to 3T. For all regions, CNR and SNR

of medulla were higher in 1.5T compared to 3T.

In the tibia, the mid diaphyseal region had higher CNR and SNR for muscles at 1.5T

than at 3T. The distal diaphyseal region had higher CNR and SNR for muscle at 3T

compared to 1.5T. In both the regions, CNR and SNR of medulla were higher at

1.5T. All of the measurement sites, with the exception of two at articular regions of

both femur and tibia, had higher CNR and SNR for soft tissue at 3T than 1.5T. A

detailed results section is available in the paper presented at the end of this chapter.


Since the introduction of MRI scanners to the clinical setting in the 1990s, a number

of studies and reviews were conducted that focused on the useability of 3T MRI

systems for scanning of various anatomical compartments, and on the various


94

advantages and disadvantages of the systems. Whilst most of these studies discussed

3T MRI scanning of the soft tissues, few studies have been conducted regarding

scanning of the skeletal system. As a radiation hazards free alternative to CT, 3T

MRI scanning for reconstruction of 3D models has become a research interest

because the higher magnetic field of 3T MRI can be used to potentially overcome

some of the limitations of the 1.5T MRI.

The present study was conducted to compare the images obtained at 1.5T and 3T,

using the instruments from the same manufacturer and identical imaging protocols.

The study resulted in higher SNR and CNR at 3T for most of the anatomical regions

compared to the 1.5T MRI. This will potentially improve the accuracy of the

articular regions of the segmented 3D models. According to the author‟s experience

of image segmentation, there is a potential for slight improvement of segmentation

time. The study had a sample size of five even though statistical power analysis

showed that a sample size of 20 is required to detect 1.76 difference between two

groups with standard deviation of 1.00. Sample size of five can detect a difference of

3.55 with the same standard deviation. Despite the small sample size, the study has

shown an overall improvement of image quality for 3T MRI compared to 1.5T MRI.

A samples size of twenty is not expected yield very different results considering that

the variability between the measurements are slightly high.

The spin relaxation times (T1 and T2) change at higher fields and the present study

showed that T1 of muscles is highly dependent on the main magnetic field and has

increased values at higher fields. T2 is relatively less dependent and takes slightly

less value at higher fields. Due to this increased T1 of muscles at 3T, theoretically,

longer TR values need to be used to obtain similar contrast to that at 1.5T and this

will, in turn, increase the scanning time. Due to this reason short TR values were

used for scanning of the human volunteers. However, for the purpose of comparison,

identical TR values were used in both 1.5T and 3T scanning protocols.

Higher magnetic fields also worsen most of the artefacts that occur in MRI scanning

of tissue compartments. The magnetic susceptibility artefact increases as higher

magnetic fields are more sensitive and this will affect the localisation of certain

anatomical structures of the images acquired. The chemical shift artefact also

becomes greatly pronounced as the frequency difference between water and fat


95

becomes double at 3T (220Hz at 1.5T and 440Hz at 3T). This increased frequency

difference, however, helps to achieve a better fat suppression of images. Increased

SAR values also should be considered with 3T MRI whereas, at 1.5T, SAR values

rarely increase beyond the limitations.

The next chapter discusses the investigation carried out to correct the step artefact

caused by volunteers moving their leg during the MRI of long bones.


96

6.8 Paper 4: 3T MRI improves bone-soft tissue image contrast

compared with 1.5T MRI (Submitted – under review)


97

3T MRI improves bone-soft tissue image contrast compared

with 1.5T MRI

Kanchana Rathnayaka1, Konstantin I Momot

2, Alan Coulthard

3, Andrew Volp

4, Tony

Sahama2, Michael A. Schütz

1,4, Beat Schmutz

1

1. Institute of Health and Biomedical Innovation, 60 Musk Avenue, Kelvin

Grove, QLD 4059, Australia

2. Queensland University of Technology, Brisbane, Australia

3. Royal Brisbane and Women‟s Hospital, Brisbane, Australia

4. Princes Alexandra Hospital, Brisbane, Australia

Submitted to Journal: Magnetic Resonance Imaging

Manuscript ID: MRI-D-11-00307

Corresponding Author:

Dr. Beat Schmutz

60 Musk Avenue

Kelvin Grove

QLD 4059, Australia

Email: [email protected]


98

Abstract

Orthopaedic implants are designed using 3D models of long bones based on accurate

computed tomography (CT), which is the gold standard for scanning of bones.

However, CT exposes a healthy human volunteer to a high dose of ionising radiation;

thus, CT is generally limited to scanning of clinical cases and cadaver specimens.

Magnetic resonance imaging (MRI), on the other hand does not involve ionising

radiation and is therefore more appropriate for scanning of healthy human volunteers

for research purposes. Current limitations of MRI include poor contrast in certain

anatomical regions and long scanning times; these limitations can potentially be

overcome by using scanners with higher field strength. This study quantitatively

compares 1.5T MRI to 3T MRI and optimises the scanning protocol of 3T MRI for a

better outcome.

Protocol optimisation was carried out by scanning the right leg of one volunteer in

three sets of images, each with varying repetition times (TR), echo times (TE) and

flip angles (FA), while keeping the other parameters constant. Longitudinal

relaxation time (T1) of muscle and bone marrow and transverse relaxation time (T2)

of muscle were calculated for both 1.5T and 3T field strengths. To compare the

images acquired at 1.5T to 3T, the right legs of five human volunteers were scanned

with 1.5T and 3T scanners from the same manufacturer (Siemens), using identical

protocols. Signal to noise ratio (SNR) and contrast to noise ratio (CNR) were

calculated for different anatomical locations of femora and tibiae.

The results show that T1 of muscle is extremely dependent on the main magnetic

field (0.9 ± 0.14 s at 1.5T and 1.5 ± 0.15 s at 3T), yielding a higher value at 3T while

T1 of bone marrow was weakly field dependent (0.25 ± 0.03 s at 1.5T and 0.30 ±0.07

s at 3T). T2 of muscle was not field dependent and was measured as 0.029 ± 0.007 s

at both 1.5T and 3T. CNR and SNR comparison of 1.5T and 3T showed a high CNR

and SNR for most regions of the femur and tibia at 3T, with the exception of the

distal diaphyseal region of the femur and the mid diaphyseal region of the tibia. The

results show that 3T MRI is expected to reduce segmentation time and potentially

will improve the accuracy of 3D models generated from such data sets compared to a

3D model generated from a 1.5T data set.


99

Introduction

Design and validation of orthopaedic implants increasingly utilises 3D models that

characterise the outer and inner geometry of long bones based on computed

tomography (CT) [1-5]. CT has become the gold standard for this purposes due to the

higher image contrast offered for the bone-soft tissue interface. CT, however,

exposes a subject to a high dose of ionising radiation thus limiting its use to scanning

of clinical cases and cadaver specimens. Due to this, imaging techniques such as

magnetic resonance imaging (MRI) which does not involve ionising radiation are

becoming more popular for scanning of long bones of healthy human volunteers for

research purposes. Some of the current limitations of using MRI for long-bone

imaging include extended scanning times and difficulty of image segmentation in

certain anatomical regions caused by the poor contrast at bone-soft tissue interfaces

of those regions [6, 7]. Higher field strength MRI scanners could potentially

overcome these limitations offering faster imaging times or better contrast levels [8].

In the present study, the gain obtained from the 3T system was solely invested in

improving the contrast at the bone-soft tissue interfaces.

Higher field strength MRI scanners (typically 3T) have been used clinically since the

1990s [9]. Since then, 3T MRI scanners have been validated for various soft tissue

compartments of the human body [10-15]. Because MRI utilises 1H nuclei as the

source of signal, it is rational to use it for studies involving soft tissues. While most

of these studies were qualitative, a few quantitative comparisons of the image quality

between 1.5T and 3T has also been reported [16]. However, to date, there are no

studies which have quantitatively compared 3T MRI with 1.5T MRI with regards to

generating 3D models of long bones.

The intrinsic signal to noise ratio (SNR) of a clinical MR system is approximately

proportional to the strength of the main magnetic field (B0) [17]. Thus, in principle, a

3T MRI system should offer twice the SNR that a 1.5T system offers if used with

equivalent parameters, receiver coils and subjects. However, the actual SNR of

acquired images is dependent on various other factors than on the doubled main

magnetic field: hardware design, change of tissue characteristics at higher fields (T1

and T2), increased sensitivity to magnetic susceptibility and the increased precession

frequency difference between fat and water. Therefore, optimisations of protocols


100

(TR, TE and FA) to meet the changed tissue characteristics and minimisation of

artefacts are important for a higher signal gain at 3T.

Longitudinal relaxation time (T1) and transverse relaxation time (T2) of a particular

tissue are two of the parameters that determine the contrast of the acquired MR

images. T1 is highly dependent on the main magnetic field (B0) and becomes longer

on increasing B0. The reported rise of T1 is about 20% for fatty tissue [10] and 40%

for muscle tissue at B0 = 3T compared to B0 = 1.5T [18]. This elongation of T1

reduces the signal intensity at shorter TRs and if similar values are used, the SNR at

3T is only slightly higher than at 1.5T. T2 is relatively less dependent on the external

magnetic field; however, about 10% reductions in T2 in certain tissue types have

been reported at 3T compared to 1.5T [18, 19]. Due to the different behaviour of spin

relaxation times at 3T, the repetition time (TR) and echo time (TE) should be

changed accordingly to obtain the maximum contrast levels. In general, a relatively

longer TR value and slightly shorter TE values should yield the optimal SNR in 3T

MR system.

SNR and contrast to noise ratio (CNR) are two most commonly used comparison

characteristics of MR images. SNR has long been used for evaluation of MR

systems, measurement of contrast enhancement, pulse sequences and RF coil

comparison [20]. CNR offers a meaningful way of comparing the contrast of the

bone-soft tissue interface of images, which is the most important feature responsible

for an accurate image segmentation, from two different field strength of MRI [21].

This study quantitatively compares the image quality at 1.5T to that at 3T using SNR

and CNR to compare the image quality. In addition, an investigation was carried out

to determine the optimum parameters to use with a FLASH (Fast Low Angle Shot)

sequence in the scanning of human volunteers.

Methods

MRI data acquisition

The MRI data for the first part of the investigation (SNR and CNR optimisation) was

acquired using 1.5T (Siemens Magnetom Avanto) and 3T (Siemens Trio Tim)

scanners. The optimisation was achieved by varying the values of one of TR, TE and

flip angle (FA), while keeping the values of the other parameters constant (Table 1).


101

The right mid femoral region of one human volunteer was scanned using a 3D

FLASH sequence. The peripheral angiography (PA) matrix coil was used to cover

the lower limb.

Table 1: The imaging protocols used to scan the human volunteer with varying TE,

TR and FA values

Parameter Varying TR Varying TE Varying FA

TR Varying 16 ms 11 ms TE 5 ms Varying 4.66 ms FA 15° 15° Varying Pixel size 0.5 mm

2 0.5 mm2 0.5 mm

2 Slice thickness 5 mm 5 mm 1 mm

Data acquisition for the second part of the investigation was carried out by scanning

the right leg of five healthy male volunteers (age range: 30-54 years) with 1.5T and

3T clinical MRI systems. A customised imaging protocol (Table 2) that was

determined in accordance with the results obtained in the first part was used for the

scanning. The legs were scanned in five segments (Figure 1), moving the table so

that the centre of each segment was positioned in the centre of the magnet. One

scanning segment contained 256 image slices and a 66 slice overlap was maintained

between two successive scanning stages.

Table 2: The MRI imaging protocols for 1.5T and 3T scanners

Parameter 1.5T 3T

In plane resolution 0.45 mm ×0.45 mm 0.45 mm ×0.45 mm Slice thickness 1 mm 1 mm TR 11 ms 11 ms TE 4.66 ms 4.66 ms Flip angle 7° 7° Number of averages 1 1 Image sequence 3D FLASH 3D FLASH Manufacturer Siemens Siemens Model Magnetom Avanto Trio Tim RF Coils PA & Body Matrix PA & Body Matrix


102

Figure 1. Five imaging segments were used to scan the lower limb completely in 3T

and 1.5T MRI scanners.

Measurement of spin relaxation times

T1 and T2* spin relaxation times of the „muscle‟ and „bone marrow‟ compartments

were measured from the appropriate series of FLASH images [22]. For the

measurement of T1, the RF excitation pulse was set to RF = 15o, the gradient echo

time to 5 ms, the number of averages to 1, and a series of TR were used. At 1.5T, the

TR values used were 10, 12, 15, 20, 30 and 50 ms. At 3T, the TR values used were

10, 14, 20, 30, 50 and 100 ms. For each voxel within the image, the T1 value was

determined from a two-parameter nonlinear least-squares fit of the intensity of the

steady-state FLASH signal as a function of TR:

(1)

Where NS = 20 was the number of slices and R1 and S0 were the adjustable fit

parameters. The fitted value of R1 was taken as the longitudinal relaxation rate, 1/T1,

in the respective voxel. The standard errors of the fitted R1 and image amplitude ( R1

and S0) were also determined for each voxel. The voxels where any of the following

conditions were observed were rejected: R1 > 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 = 15

o, 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,

s

S

NRTR

RF

NRTR

e

eSTRS

1

1

cos1

1)( 0


103

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:

(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).

(3)

(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 [20] was introduced to standardise the SNR and CNR

*2

0)(RTE

eSTES

Noise

SignalSNR

Noise

SignalSignalCNR 21

)4/(2


104

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:

(5)

(6)

Where, SNR = signal to noise ratio, CNR = contrast to noise ratio, Mtissue = 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.

bone

tissue

STD

MSNR

4

2

bone

tissuetissue

STD

MMCNR

4

2

21


105

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.


106

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, e-

proximal 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.


107

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.


108

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).


109

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).


110

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).


111

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).


112

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


113

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


114

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


115

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, bone-

tendon, 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.


116

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):265-

275.

[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-to-

noise 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.


117

[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.


118

[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.


119

[27]. Cavanagh J, Fairbrother WJ, Palmer AG, Rance M, Skelton NJ. Protein NMR

Spectroscopy: Principles and Practice. 2nd ed. 2007, Burlington: Elsevier

Academic Press.

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

121

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


122

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).


123

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.


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


124

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.


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


125

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.


126

7.7 Paper 5: Correction of step artefact associated with MRI

scanning of long bones (Submitted – under review)


127

Correction of step artefact associated with MRI scanning of

long bones

1Kanchana Rathnayaka,

2Gary Cowin,

1,3Michael A Schuetz,

4Tony Sahama,

1Beat

Schmutz

1Institute of Health and Biomedical Innovation, Brisbane, QLD, Australia

2University of Queensland, St Lucia, QLD, Australia

3Princess Alexandra Hospital Brisbane, QLD, Australia

4Queensland 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]


128

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.


129

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


130

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.


131

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 Value

Instrument Siemens Trio tim

Field Strength 3T

In plane resolution 0.47 mm × 0.47 mm

Slice thickness 1 mm

TE 1.83 ms

TR 11 ms

FA 10°

Image sequence 3D VIBE

Number of Averages 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.


132

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 multi-

threshold 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


133

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


134

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,


135

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

3D models.

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

0.00

0.02

0.04

0.06

0.08

0.10

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

Aver

age

dev

iati

on

(m

m)

Overlap (cm)


136

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.


137

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

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

No table movement With table movement

With corrected lateral shift artefact

De

via

tio

n ±

SD

(m

m)


138

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

reference models.

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

Axis X -X

Initial deviation (mm) 5 10 15 5 10 15

Maximum (mm) 2.54831 2.55382 2.55255 2.54727 2.55275 2.51663 Average (mm) 0.31716 0.31729 0.31725 0.31713 0.31722 0.31662

SD 0.25116 0.25187 0.25164 0.25099 0.25172 0.24708

Axis Y -Y

Initial deviation (mm) 5 10 15 5 10 15

Maximum (mm) 2.54655 2.54958 2.55511 2.54944 2.54584 2.54726

Average (mm) 0.31708 0.31719 0.31733 0.31715 0.31710 0.31714 SD 0.25094 0.25134 0.25201 0.25135 0.25081 0.25101

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

No table movement

With table movement

With corrected lateral shift

artefact

Single scan MRI

De

via

tio

n ±

SD (

mm

)

* *

*

*


139

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


140

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.


141

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.


142

[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):179-

191.

[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):1401-

1411.

[11]. Besl PJ, McKay ND. A Method for Registration of 3-D Shapes. IEEE

Transaction on Pattern Analysis and Machine Intelligence 1992;14(2):239-

256.

[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.


143

[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.

Chapter 8: Summary, conclusion and future directions

145

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


146

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


147

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 (T1) 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


148

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


149

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


150

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.

Appendix 1

151

Appendix 1 Ethical approval for the study in Chapter 6

••

-,

~

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

Queensland Government

Office of the Human Research Ethics Committees Queensland Health

Or Kanchana Rathnayaka Queensland University ofTechnology Institute of Health & Biomedical Innovation 60 Musk A venue Kelvin Grove Q 4059

Dear Dr Rathnayaka,

Em1uioics 10: Odcnc Pctco,;cn Coordinator

Phone: 07 3636 5490 Fax: 07 3636 5849 Our Rcf: HREC/1 0/QRBW/1 41 E-mail IW\1'11-I'oluc''" healt h qld go,·.au

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.</1tmc & Women·_. /ln;pital /lunwn Rcst•llrcil Ethic.< Commillct• is cvmlilulcd ami operate.< according lo tile NNMRC's Nmioual Swu·mcnl ou l:'thical Couducl ill /fumau Rc.<carch (2 1!117).

Office

Bullcrlidd St reet l·krstun Q ~02'1

Postal

Pnst Ollicc Hcrstun Queensland 4029 Australia

Phone Fnx

07 3636 541)0 07 3636 5849 lSD + 61 7 3636 541JO

Appendix 1

152

ii•J,ro/ /lri.rholl<' & Woml'll '.< llmpilol /111/o'C' 1/cf N": 1//II:"Ci/ INJIIIJII'I / .f I

Document

Covering Letter

Application: NEAF

Protocol: MR protocol for scanning tibia & femur of a volunteer

Curriculum Vitae of Kanchana Ratl~nay:~~a

Version .,l

I 2.0 (2008)

/3.115.111/11

Date -·-_I

29 March 20 I 0 I i ................. .. I I I March 2010 I

-' --- -~ --- _.. -· ___ _j . --··· -· ···- _,_, _________ - _j

Curriculum Vitae of Beat Schmutz . i __ --· --·. _! ------·-------·---: ' Letter of Support from Professor Alan Coulthard, Dept of I '1 23 March 2010 !

M d. 11 . I I e 1ca magmg 1 ... J. . .......... ___ __I

Research Funding Schedule (reviewed in accordance with I l _j Section 3.3. /§_o[!lze _Nc:J..~ona~ Sta_f_!!!J.c:!!JL .. --·---.. _ ______ ]_______ ·--

Emai I coJTespo_n~en<:e ~~ Q!:!I l_n~~m~~ie~. ~EJ!:l~.':l-~nc~---·_j ___ __] 1 0 Ma~J_O __ j Response to R_~q!:l~tf'?!. _F!.!l'!!~e~In_fo~nation ____________ .. ___ __j ______ j _j_!.Jv.1~Y- 20lQ__ j

Participant lnfOTJTI~~ion Sl~e~-~-~O':JS_el_l! ..f.<?!.!'!!_ ....... _____ j __ 2 ____ _ __1 _I_!_~~ :X: 20 1_0 __ j

Please note the following conditions of approval:

I. The P1incipal Investigator will immediately rep011 anything which might wrunnt review of ethical approval of the project in the specified fon'nat, including:

• Unforeseen events that might affect continued ethical acceptability of the project. Serious Adverse Events must be notified to the Committee as soon as possible. In addition, the Investigator must provide a summary of the adverse events, in the specified fmmat, including a comment as to suspected causality and whetl1er changes are required to the Patient lnfonnation and Consent Form. In the case of SeJious Adverse Events occuJTing at the local site, a full report is required from the Principal Investigator, including duration of treatment and outcome of event.

2. Amendments which do not affect either the ethical acceptability or site acceptability of the project (e.g. typographical eJTors) should be submitted in hard copy to the HREC Coordinator. These should include a covering letter from the Principal Investigator providing a brief descliption of the changes and the rationale for the changes, and accompanied by all relevant updated documents with tracked changes.

3. Proposed amendments to the research project which may affect both the ethical acceptability and site suitability of the project must be submitted firstly to the HREC for review and, once HREC approval has been granted, then submitted to the Research Governance Office.

4. Amendments to the research project which only affect the ongoing site acceptability of the project are not required to be submitted to the HREC for review. These amendment requests should be submitted directly to the Research Governance Office

Appendix 1

153

lio!l a/ 1/ri ,·hall<' & ll'omwn 's 1/ospillll /III/X ' llr:{-No: 1/1/I:'C/ /11/Q/IIl ll'/1·11

(by-passing the HREC).

3 IJ.05.211/IJ

5. Amendments to the research project which may affect the ongoing ethical acceptability of a project must be submitted to the HREC for review. Major amendments should be reflected in a revised online NEAF (accompanied by all relevant updated documentation and a coveting letter fi·om the Principal Investigator, providing a brief desctiption of the changes, the rationale for the changes, and their implications for the ongoing conduct of the study). Hard copies of the revised NEAF, the cover letter and all relevant updated documents with tracked changes must also be submitted to the HREC Coordinator as per standard HREC SOP. Further advice on submitting amendments is available fi·om hllp: . \\' lV II' .hc:1l th .qld. !!.I >v .au/uhnl r/dlH.:umcnls/n:scarchcr uscr!!.uidc.pd f

6. The I-IREC will be notified, giving reasons, if the project is discontinued at a site before the expected date of completion.

7. The HREC will be notified, giving reasons, on any sponsor repmis or other infonnation which might affect the ongoing ethical acceptability in line with the requirements of the ICH GCP guidelines as annotated by the TGA: hll p :ti ii'WW .ll!a.l!OV .au/doc.:s/pd li'CUl!Uidc/ich/ich [J 51J5.pd f

8. The Ptincipal Investigator will provide an Annual Repoti to the HREC and at completion of the study in the specified fonnat.

9. The District Administration and the Human Research Ethics Committee may inquire into the conduct of any research or purpotied resem'ch, whether approved or not and regardless of the source of funding, being conducted on Hospital premises or claiming any association with the Hospital, or which the Committee has approved if conducted outside Royal Brisbane & Women's Hospital Metro Nmih Health Service District.

Should you have any queries about the HREC's consideration of your project please contact I lie Ill\ I:C ('pnrtfinalor on 07 363(> 5.:JlJO. The HREC terms of Reference, Standard Operating Procedures, membership and standard forms are available from li lip :i w 11 11· . l1 L'a it h. ql d .l!O v .J u.'nl11 nr/h t 1111 /n:l!ui rcgu hnmc.asp

Once authorisation to conduct the research has been granted, please complete the Commencement Fom1 ( : I twcluncnt 1/J and retum to the office ofthe Human Research Ethics Committee.

The HREC wishes you every success in your research.

Yours sincerely, -1 . --,J

' / ,, -'~r . --, -- ')

. ~

Or Conor Brophy Chair·pcrson RBWH Human Research Ethics Committee ,\ ·ki r•' N, >rl h District 13.05.2010

Appendix 2

154

Appendix 2 Participant information and Consent form

used in Chapter 6

~ ~n~u~fi Health and Biomedical Innovation

Queensland Government Queensland Health

PARTI:PANT lNFORMATDN for RESEARCH PROJECT

"Comparative study of 3T M RI vs l.ST for the acquisition of 3D mor holo ical bone data of the lower extremi "

Research Team Contacts Dr Kanchana Rathnayaka, Principal Investigator

Institute of Health and Biomedical Innovation F acuity of Built Environment and Engineering

(07) 3138 6234 k.rathnava ka@ aut.edu .au

Description

Dr. B. Schmutz, Senior Research Fellow Institute of Health and Biomedical Innovation Faculty of Built Environment and Engineering

(07) 3138 6238 b.schmutz@ aut.edu.au

This research is undertaken as a part of the PhD project of Kanchana Rathnayaka, Trauma Research Group (www.ihbi.gut.edu.au/qo/mdtrauma) at the Queensland University of Technology (QUT), with the image data being acquired at the Royal Brisbane and Women's Hospital.

The purpose of this project is to assess and quantify the improvement in image quality obtained with a high field 3 Tesla (3T) Magnetic Resonance Imaging (M RI) scanner compared to images acquired with a 1.5 Tesla (l.ST) conventional scanner. Some of the current limitations of MRI are long scanning times and low image contrast for certain anatomical regions. Higher field strength (3T and above) scanners offer improved signal which may be translated to faster imaging times or better image quality.

The research team requests your assistance because you have previously participated in a similar project (Human bone morphology database (MRI of volunteers) where l.ST M RI images of your leg were acquired. The images of your leg, acquired in the previous study, can now be compared with those acquired with a stronger 3T magnet in this study.

Participation Your participation in this project is voluntary. If you do agree to participate, you can withdraw from participation at any time before your data has been acquired without comment or penalty. Your decision to participate will in no way impact upon your current or future relationship with QUT, with the Royal Brisbane and Women's Hospital or with Queensland Health.

Your participation will involve the acquisition of anonymous (non-identifiable) Magnetic Resonance Image (MRI) data of your legs. You will also be asked to provide details of your age, gender, height and weight which will be stored with your image data.

It is expected that it will take approximately 45- 60 minutes to acquire the image data of your leg, during that time you will be asked not to move your leg which is being imaged. To reduce the level of noise generated during the operation of the M RI scanner you will be given earplugs or headphones to wear (you may bring along your choice of relaxing music on a CD). Your total stay in the Radiology Department could last up to 2 hours.

The imaging session will be conducted at the Royal Brisbane and Women's Hospital, Butterfield Street, Herston, QLD 4029, Australia. The data will be securely stored and analysed at QUT.

You will receive Coles-Myer's Gift Vouchers equal to an amount of $50.- as reimbursement for your travel expenses and for your time spent in association with this project.

Expected benefits It is expected that this project will not benefit you. The outcome if the project will facilitate the generation of 3D models from 3T M RI images using less segmentation time compared to the l.ST scanner images.

Volunteer info consentfonn Version 3. 8 June 2010 1 of 3

Appendix 2

155

• ~n~uefi Health and Biomedical Innovation

Risks


There are no risks beyond normal day-to-day living associated with your participation in this project. Because M RI uses low-energy, non-ionising radio waves, there are no known risks or side effects except for the groups specified below.

While there are no known hazards, MRI is not proven to be safe during the first three months of pregnancy. Therefore if you are pregnant (or if you suspect that you are pregnant) we ask you not to participate in this study.

The magnet at the centre of the scanner may affect, or be affected by, any person fitted with a pacemaker, hearing aid, other electrical device, or metal implants. If you are fitted with any such devices or implants we ask you not to participate in this study.

You wil l be asked not to move your leg which is being imaged for approximately 45 - 60 minutes during the scanning. If you have any medical conditions which are affected by this, we ask you not to participate in this study.

If you have sustained a penetrating eye injury in the past, there might be harmful effects from M RI scanning and we ask you not to participate in this study.

For a part of the imaging session your whole body will be inside the confined space of the scanner's tunnel. Therefore, if you are claustrophobic we ask you not to participate in this study.

Confidentiality All comments and responses are anonymous and will be treated confidentially. The names of individual persons are not required in any of the responses. Your name will not be recorded or stored with your image data.

Your data will be stored on a secure server at QUT. Access to the database is not available to the public. Upon request, the Trauma Research Group atQUT reserves the right to share your data with third parties to be used in projects aiming to benefit human kind.

Thus, the data acquired from your leg might be used for projects which involve research, development, education and teaching.

Consent to Participate We would like to ask you to sign a written consent form (enclosed) to confirm your agreement to participate.

Questions /further information about the project Please contact the research team members named above to have any questions answered or if you require further information about the project.

Concerns /complaints regarding the conduct of the project QUT and the Royal Brisbane and Women's Hospital are committed to researcher integrity and the ethical conduct of research projects. However, if you do have any concerns or complaints about the ethical conduct of the project you may contact the Human Research Ethics Committee (HREC) Coordinator at the Royal Brisbane and Women's Hospital on (07) 3636 5490 and/or Research Ethics Coordinator at Queensland University of Technology on (07) 31382091. The HREC Coordinator is not connected with the research project and can facilitate a resolution to your concern in an impartial manner.

Volunteer info consent form Version 3, 8 June 2010 2 of 3

Appendix 2

156

• ~n~u~fi Health and Biomedical lnnoval ion


CONSENT FORM for RESEARCH PROJECT

"Comparative study of 3T M RI vs l.ST for the acquisition of 30 mor holo ical bone data of the lower extremi "

Statement of consent

By signing below, you are indicating that you:

• have read and understood the information document regarding this project

• have had any questions answered to your satisfaction

• have been given the opportunity to have a friend or relative present when the study was explained

• understand that if you have any additional questions you can contact the research team

• understand that you are free to withdraw at any time before your data has been entered into the database, without comment or pena lty

• understand that you can contact the HREC Coordinator at the Royal Brisbane and Women's Hospital on (07) 3636 5490 if you have concerns about the ethic a I conduct of the project

agree to participate in the project which involves the acquisition and storage of anonymous (nonidentifiable) magnetic resonance image data of your leg

• agree that the data which has been acquired of your leg can be used for various projects in research, development, education and teaching

• agree that images of your bones might be used for publications, teaching and educationa l purposes

• are not pregnant to the best of your knowledge

• have not sustained any penetrating eye injuries in the past

Participant Name ____________________________ _

Signature ----·---------·--·--·----------------------·---------------------------·--·-------·--· -·-------·----·------·---------------

Date __ _ I I

I have explained the nature and purpose of this study to the above participant and have answered their questions.

~vestigarorName ___________________________ _

Signature __ _

Date I I

Volunteer info consent form Version 3, 8 June 2010 3 of 3

Appendix 3

157

Appendix 3 Animal tissue use notification

Dear Dr Kanchana Rathnayaka Mudiyanselage,

Re: TISSUE USE NOTIFICATION: Use of ovine limbs for a study entitled

"Correction of the step artefact associated with MRI of long bones" (Source

studies: 08-0848, Goss)

Thank you for your notification of animal tissue use which has been noted and

confirmed as falling outside the scope of requiring review by the UAEC.

Your confirmed application is attached and your approval number is:

1000000529. Please quote this number in all future correspondence.

SPECIFIC CONDITIONS OF APPROVAL:

a. the provision of specimen material is approved by the CI of the source study (i.e.

Dr Ben Goss) b. there are no changes to the source study's protocol to facilitate

provision of the specimens:

- the animal is only killed by the person authorised to do so

- the researcher has no input into the treatment and handling of the

animal prior to euthanasia

- the researcher has no input into the timing or manner of euthanasia

- the tissue/whole animal is collected by the researcher after death is

confirmed.

Note: Tissue use notifications are made available to the UAEC for noting at the next

available meeting. You will only be contacted again if the UAEC raise any questions

at that time.

Please do not hesitate to contact the Research Ethics Unit if you have any queries.

Best regards

Research Ethics Unit | Office of Research | Level 4 | 88 Musk Ave | Kelvin Grove

p: +61 7 3138 2340 | f: +61 7 3138 4543 | e: [email protected]

w: http://www.research.qut.edu.au/ethics/

mailto:[email protected]

http://www.research.qut.edu.au/ethics/

References

159

References

1. Lee, Y., et al., Anatomical evaluation of CT-MRI combined femoral model.

Biomedical Engineering Online, 2008. 7(1): p. 6.

2. Gelaude, F., J.V. Sloten, and B. Lauwers, Semi-automated segmentation and

visualization of outer bone cortex from medical images. Computer Methods

in Biomechanics & Biomedical Engineering, 2006. 9(1): p. 65-77.

3. Gelaude, F., J.V. Sloten, and B. Lauwers, Accuracy assessment of CT-based

outer surface femur meshes. Computer Aided Surgery, 2008. 13(4): p. 188-

199.

4. Gassman, E.E., et al., Automated bony region identification using artificial

neural networks: reliability and validation measurements. Skeletal

Radiology, 2008. 37: p. 313-319.

5. Pilgram, R., et al., Knowledge-based femur detection in conventional

radiographs of the pelvis. Computers in Biology and Medicine, 2008. 38(5):

p. 535-44.

6. Revol-Muller, C., et al., Automated 3D region growing algorithm based on

an assessment function. Pattern Recognition Letters, 2002. 23(1-3): p. 137-

150.

7. Van Luong, H. and J. Kim, A New Parallel Approach to Fuzzy Clustering for

Medical Image Segmentation, in Advances in Visual Computing. 2008. p.

1092-1101.

8. Noble, P.C., et al., The effect of aging on the shape of the proximal femur.

Clin Orthop Relat Res, 1995(316): p. 31-44.

9. Henley, G. and J.E. Harrison, 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.

References

160

10. Cao, H., et al., Quantitative 31P NMR spectroscopy and 1H MRI

measurements of bone mineral and matrix density differentiate metabolic

bone diseases in rat models. Bone, 2010. 46(6): p. 1582-1590.

11. Gold, G.E., et al., MR Spectroscopic imaging of collagen: Tendons and knee

menisci. Magnetic Resonance in Medicine, 1995. 34(5): p. 647-654.

12. Reichert, I.L.H., et al., Magnetic resonance imaging of cortical bone with

ultrashort TE pulse sequences. Magnetic Resonance Imaging, 2005. 23(5): p.

611-618.

13. Moore, S., et al., 3D models of blood flow in the cerebral vasculature. Journal

of Biomechanics, 2006. 39(8): p. 1454-1463.

14. Lösch, A., et al., A non-invasive technique for 3-dimensional assessment of

articular cartilage thickness based on MRI part 1: Development of a

computational method. Magnetic Resonance Imaging, 1997. 15(7): p. 795-

804.

15. Haubner, M., et al., A non-invasive technique for 3-dimensional assessment of

articular cartilage thickness based on MRI part 2: Validation using CT

arthrography. Magnetic Resonance Imaging, 1997. 15(7): p. 805-813.

16. Felix Eckstein, D.B.T.M.L., Quantitative MRI of cartilage and bone:

degenerative changes in osteoarthritis. NMR in Biomedicine, 2006. 19(7): p.

822-854.

17. Augat, P. and F. Eckstein, Quantitative Imaging of Musculoskeletal Tissue.

Annual Review of Biomedical Engineering, 2008. 10(1): p. 369-390.

18. Schmidt, G.P., et al., Whole-body MR imaging of bone marrow. European

Journal Of Radiology, 2005. 55(1): p. 33-40.

19. Hoad, C.L., et al. Development of a 3D MRI sequence for computer assisted

surgery of the lumbar spine. in ISMRM. 8th Scientific Meeting and Exhibition

(Denver, Colorado). 2000: Proc. Intl. Soc. Mag. Reson. Med.

References

161

20. Martel, A.L., et al., Assessment of 3-dimensional magnetic resonance

imaging fast low angle shot images for computer assisted spinal surgery.

Computer Aided Surgery, 1998. 3(1): p. 40-44.

21. Fassbind, M., et al. Quantifying foot bone motion via magnetic resonance

imaging. in 52nd Annual Meeting of the Orthopaedic Research Society. 2006.

Chicago, IL: The orthopaedic research society.

22. Edelstein, W.A., et al., The intrinsic signal-to-noise ratio in NMR imaging.

Magnetic Resonance in Medicine, 1986. 3(4): p. 604-618.

23. Ramnath, R.R., 3T MR imaging of the musculoskeletal system (Part I):

considerations, coils, and challenges. Magnetic Resonance Imaging Clinics

of North America, 2006. 14(1): p. 27-40.

24. Hoad, C.L., et al., A 3D MRI sequence for computer assisted surgery of the

lumbar spine. Physics in Medicine and Biology, 2001. 46: p. N213-N220.

25. Wolf, P., et al., Reliability of tarsal bone segmentation and its contribution to

MR kinematic analysis methods. Computerized Medical Imaging and

Graphics, 2007. 31(7): p. 523-530.

26. Schmutz, B., et al., Using MRI for the imaging of long bones: First

Experience. Journal of biomechanics, 2008. 41(Supplement1): p. S188.

27. Cogbill, T.H. and K.J. Zeigelbein, Computed Tomography, Magnetic

Resonance, and Ultrasound Imaging: Basic Principles, Glossary of terms,

and Patient Safety. Surgical Clinics of North America, 2011. 91(1): p. 1-14.

28. Novelline, R.A., et al., Helical CT in Emergency Radiology1. Radiology,

1999. 213(2): p. 321-339.

29. Blackmore, C.C., et al., Cervical Spine Screening with CT in Trauma

Patients: A Cost-effectiveness Analysis1. Radiology, 1999. 212(1): p. 117-

125.

30. HALL, E.J. and D.J. BRENNER, Cancer risks from diagnostic radiology.

British Journal of Radiology, 2008. 81(965): p. 362-378.

References

162

31. Smith-Bindman, R., et al., Radiation Dose Associated With Common

Computed Tomography Examinations and the Associated Lifetime

Attributable Risk of Cancer. Archives of Internal Medicine, 2009. 169(22): p.

2078-2086.

32. Berrington de Gonzalez, A., et al., Projected Cancer Risks From Computed

Tomographic Scans Performed in the United States in 2007. Archives of

Internal Medicine, 2009. 169(22): p. 2071-2077.

33. Griffey, R.T. and A. Sodickson, Cumulative Radiation Exposure and Cancer

Risk Estimates in Emergency Department Patients Undergoing Repeat or

Multiple CT. American Journal of Roentgenology, 2009. 192(4): p. 887-892.

34. Raeburn, S., Radiation doses in computed tomography: The increasing doses

of radiation need to be controlled. BMJ, 2000. 320: p. 593-594.

35. Wall, B. and D. Hart, Revised radiation doses for typical X-ray examinations.

Report on a recent review of doses to patients from medical X-ray

examinations in the UK by NRPB. National Radiological Protection Board.

British Journal of Radiology, 1997. 70(833): p. 437-439.

36. Radiologists, T.R.C.o., Making the best use of department of clinical

radiology: guidelines for doctors. 5th ed. 2003, London: Royal College of

Radiologists.

37. Semelka, R.C., et al., Imaging strategies to reduce the risk of radiation in CT

studies, including selective substitution with MRI. Journal of Magnetic

Resonance Imaging, 2007. 25(5): p. 900-909.

38. Zannoni, C., A. Cappello, and M. Viceconti, Optimal CT scanning plan for

long-bone 3-D reconstruction. Medical Imaging, IEEE Transactions on,

1998. 17(4): p. 663-666.

39. Sint Jan, S.V., et al., Low-dose computed tomography: A solution for in vivo

medical imaging and accurate patient-specific 3D bone modeling? Clinical

Biomechanics, 2006. 21(9): p. 992-998.

References

163

40. Brown, M.A. and R.C. Semelka, MRI Basic principles and applications. 4th

ed. 2010, New Jersey: John Wiley & Sons.

41. Lipton, M.L., Totally Accessible MRI. 2008, New York: Springer.

42. Kangarlu, A. and P.-M.L. Robitaille, Biological effects and health

implications in magnetic resonance imaging. Concepts in Magnetic

Resonance, 2000. 12(5): p. 321-359.

43. Merkle, E.M. and B.M. Dale, Abdominal MRI at 3.0 T: The Basics Revisited.

American Journal of Roentgenology, 2006. 186(6): p. 1524-1532.

44. Erasmus, L., et al., A short overview of MRI artefacts. South African Journal

of Radiology, 2004. 8(2): p. 13-17.

45. Schmutz, B., et al., Fit Assessment of Anatomic Plates for the Distal Medial

Tibia. Journal of Orthopaedic Trauma, 2008. 22(4): p. 258-263.

46. Goyal, K.S., et al., Analysis of anatomic periarticular tibial plate fit on

normal adults. Clinical Orthopaedics and Related Research, 2007. 461: p.

245-257.

47. Schmutz, B., et al., Fit optimisation of a distal medial tibia plate. Computer

Methods in Biomechanics & Biomedical Engineering, 2011. 14(4): p. 359-

364.

48. Park, A., J. DiStefano, and T. Nguyen, 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.

49. Schmutz, B., et al., Quantitative fit assessment of tibial nail designs using 3D

computer modelling. Injury, 2010. 41(2): p. 216-219.

50. Wehner, T., Simulation der Frakturheilung beim Fixateur Externe und bei

intramedullarem Nagel., in Medizinische Fakultat. 2008, Universitat Ulm.

References

164

51. Niu, Q., et al., Image processing, geometric modeling and data management

for development of virtual bone surgery system. Computer Aided Surgery,

2008. 13(1): p. 30-40.

52. Handels, H., et al., Virtual planning of hip operations and individual

adaption of endoprostheses in orthopaedic surgery. International Journal of

Medical Informatics, 2000. 58-59: p. 21-28.

53. Zachow, S., et al., Draw and cut: intuitive 3D osteotomy planning on

polygonal bone models. International Congress Series, 2003. 1256: p. 362-

369.

54. Xia, J., et al., Computer-assisted three-dimensional surgical planning and

simulation: 3D color facial model generation. International Journal of Oral

and Maxillofacial Surgery, 2000. 29(1): p. 2-10.

55. Gomez-Benito, M.J., et al., A 3D Computational Simulation of Fracture

Callus Formation: Influence of the Stiffness of the External Fixator. Journal

of Biomechanical Engineering, 2006. 128(3): p. 290-299.

56. Lacroix, D. and P.J. Prendergast, Three-dimensional simulation of fracture

repair in the human tibia. Computer Methods in Biomechanics and

Biomedical Engineering, 2002. 5(5): p. 369-76.

57. Anwar, A., A. Mekonnen, and P. Diamantopoulos, FEA modelling of

anatomical structures. Journal of Biomechanics, 2006. 39(Supplement 1): p.

S423-S423.

58. Trabelsi, N., Z. Yosibash, and C. Milgrom, Validation of subject-specific

automated p-FE analysis of the proximal femur. Journal of Biomechanics,

2009. 42(3): p. 234-241.

59. Keyak, J.H. and Y. Falkinstein, Comparison of in situ and in vitro CT scan-

based finite element model predictions of proximal femoral fracture load.

Medical Engineering & Physics, 2003. 25(9): p. 781-787.

References

165

60. Pérez del Palomar, A. and M. Doblaré, Dynamic 3D FE modelling of the

human temporomandibular joint during whiplash. Medical Engineering &

Physics, 2008. 30(6): p. 700-709.

61. MacNeil, J.A. and S.K. Boyd, Bone strength at the distal radius can be

estimated from high-resolution peripheral quantitative computed tomography

and the finite element method. Bone, 2008. 42(6): p. 1203-1213.

62. Viceconti, M., et al., CT data sets surface extraction for biomechanical

modeling of long bones. Computer Methods and Programs in Biomedicine,

1999. 59(3): p. 159-166.

63. Silva, M.J., et al., Direct and computed tomography thickness measurements

of the human, lumbar vertebral shell and endplate. Bone, 1994. 15(4): p. 409-

414.

64. Dougherty, G. and D. Newman, Measurement of thickness and density of thin

structures by computed tomography: A simulation study. Medical Physics,

1999. 26(7): p. 1341-1348.

65. Maloul, A., J. Fialkov, and C. Whyne, The Impact of Voxel Size-Based

Inaccuracies on the Mechanical Behavior of Thin Bone Structures. Annals of


66. Prevrhal S, et al., Accuracy of CT-based thickness measurement of thin

structures: modeling of limited spatial resolution in all three dimensions.

Medical Physics, 2003. 30(1): p. 1-8.

67. Souza, A., J.K. Udupa, and P.K. Saha, Volume rendering in the presence of

partial volume effects. Medical Imaging, IEEE Transactions on, 2005. 24(2):

p. 223-235.

68. Heuscher, D.J. and M. Vembar, Reduced partial volume artifacts using spiral

computed tomography and an integrating interpolator. Medical Physics,

1999. 26(2): p. 276-286.

References

166

69. Schmutz, B., M.E. Wullschleger, and M.A. Schuetz. The effect of CT slice

spacing on the geometry of 3D models. in Proceedings 6th Australasian

Biomechanics Conference. 2007. Auckland, New Zealand: The University of

Auckland.

70. Drapikowski, P., Surface modeling-Uncertainty estimation and visualization.

Computerized Medical Imaging and Graphics, 2008. 32(2): p. 134-139.

71. Kurmis, A.P., J.P. Slavotinek, and K.J. Reynolds, The influence of slice

thickness on the volume measurement accuracy of 3-D MR reconstructions of

acrylic phantoms: a precursor to knee imaging. Radiography, 2004. 10(4): p.

277-285.

72. Pal, N.R. and S.K. Pal, A review on image segmentation techniques. Pattern

Recognition, 1993. 26(9): p. 1277-1294.

73. Olabarriaga, S.D. and A.W.M. Smeulders, Interaction in the segmentation of

medical images: A survey. Medical Image Analysis, 2001. 5(2): p. 127-142.

74. Pham, D.L., C. Xu, and J.L. Prince, Current methods in medical image

segmentation. Annual Review of Biomedical Engineering, 2000. 2(1): p. 315-

337.

75. Behiels, G., et al., Evaluation of image features and search strategies for

segmentation of bone structures in radiographs using Active Shape Models.

Medical Image Analysis, 2002. 6(1): p. 47-62.

76. Lee, H., et al., Left Ventricular segmentation using graph searching on

intensity and gradiant and a priori knowledge (1vGIGA) for short-axis

cardiac mgnetic resonance imaging. Journal of Magnetic Resonance

Imaging, 2008. 28: p. 1393-1401.

77. Santarelli, M.F., et al., Automated cardiac MR image segmentation: theory

and measurement evaluation. Medical Engineering & Physics, 2003. 25(2): p.

149-159.

References

167

78. Zhou, X., et al., Automated segmentation of mammary gland regions in non-

contrast X-ray CT images. Computerized Medical Imaging and Graphics,

2008. 32(8): p. 699-709.

79. Zoroofi, R.A., et al., Automated segmentation of necrotic femoral head from

3D MR data. Computerized Medical Imaging and Graphics, 2004. 28(5): p.

267-278.

80. Peter, Z., et al., A constrained region growing approach based on watershed

for the segmentation of low contrast structures in bone micro-CT images.

Pattern Recognition, 2008. 41(7): p. 2358-2368.

81. Hirano, S. and Y. Hata, Fuzzy expert system for foot CT image segmentation.

Image and Vision Computing, 2001. 19(4): p. 207-216.

82. Cristoforetti, A., et al., Isolation of the left atrial surface from cardiac multi-

detector CT images based on marker controlled watershed segmentation.

Medical Engineering & Physics, 2008. 30(1): p. 48-58.

83. DeVries, N., et al., Validation of phalanx bone three-dimensional surface

segmentation from computed tomography images using laser scanning.

Skeletal Radiology, 2008. 37(1): p. 35-42.

84. Haas, B., et al., Automatic segmentation of thoracic and pelvic CT images for

radiotherapy planning using implicit anatomic knowledge and organ-specific

segmentation strategies. Physics in Medicine and Biology, 2008. 53(6): p.

1751-1771.

85. de Bruin, P.W., et al., Interactive 3D segmentation using connected

orthogonal contours. Computers in Biology and Medicine, 2005. 35(4): p.

329-346.

86. Bokde, A.L.W., et al., Reliable manual segmentation of the frontal, parietal,

temporal, and occipital lobes on magnetic resonance images of healthy

subjects. Brain Research Protocols, 2005. 14(3): p. 135-145.

References

168

87. Stammberger, T., et al., Interobserver reproducibility of quantitative

cartilage measurements: comparison of B-spline snakes and manual

segmentation. Magnetic Resonance Imaging, 1999. 17(7): p. 1033-1042.

88. Bryce, C.D., et al., Validation of three-dimensional models of in situ

scapulae. Journal of Shoulder and Elbow Surgery, 2008. 17(5): p. 825-832.

89. Ward, K.A., J.E. Adams, and T.N. Hangartner, Recommendations for

Thresholds for Cortical Bone Geometry and Density Measurement by

Peripheral Quantitative Computed Tomography. Calcified Tissue

International, 2005. 77(5): p. 275-280.

90. Jong Bae, K., et al., A Real-Time Region-Based Motion Segmentation Using

Adaptive Thresholding and K-Means Clustering. Lecture notes in computer

science, 2001. 2256: p. 213-224.

91. Wiemker, R. and V. Pekar, Fast computation of isosurface contour spectra

for volume visualization. International Congress Series, 2001. 1230: p. 389-

394.

92. Batenburg, K.J. and J. Sijbers, Optimal Threshold Selection for Tomogram

Segmentation by Projection Distance Minimization. Medical Imaging, IEEE

Transactions on, 2009. 28(5): p. 676-686.

93. Guo, R. and S.M. Pandit, Automatic threshold selection based on histogram

modes and a discriminant criterion. Machine Vision and Applications, 1998.

10(5): p. 331-338.

94. Wang, L. and J. Bai, Threshold selection by clustering gray levels of

boundary. Pattern Recognition Letters, 2003. 24(12): p. 1983-1999.

95. Zhang, J., et al., Fast segmentation of bone in CT images using 3D adaptive

thresholding. Computers in Biology and Medicine, 2010. 40(2): p. 231-236.

96. Marr, D. and E. Hildreth, Theory of Edge Detection. Proceedings of the Royal

Society of London. Series B. Biological Sciences, 1980. 207(1167): p. 187-

217.

References

169

97. Testi, D., et al., Border-tracing algorithm implementation for the femoral

geometry reconstruction. Computer Methods and Programs in Biomedicine,

2001. 65(3): p. 175-182.

98. Canny, J., A computational approach to edge detection. IEEE Transaction on

Pattern Analysis and Machine Intelligence, 1986. 8(6): p. 679-698.

99. Demigny, D., A discrete expression of Canny's criteria for step edge detector

performances evaluation. 1997. p. 1199-1211.

100. Bourennane, E., et al., Generalization of Canny-Deriche filter for detection of

noisy exponential edge. Signal Processing, 2002. 82(10): p. 1317-1328.

101. Heath, M.D., et al., A robust visual method for assessing the relative

performance of edge-detection algorithms. Pattern Analysis and Machine

Intelligence, IEEE Transactions on, 1997. 19(12): p. 1338-1359.

102. Adams, R. and L. Bischof, Seeded region growing. Pattern Analysis and

Machine Intelligence, IEEE Transactions on, 1994. 16(6): p. 641-647.

103. Yan, K., K. Engelke, and W.A. Kalender, A new accurate and precise 3-D

segmentation method for skeletal structures in volumetric CT data. Medical

Imaging, IEEE Transactions on, 2003. 22(5): p. 586-598.

104. Syoji, K., K. Katsuya, and H. Yutaka, Fully Automated Segmentation of

Cerebral Ventricles from 3-D SPGR MR Images using Fuzzy Representative

Line. Soft computing, 2006. 10.

105. Ahmadian, A., et al., Correction of oral contrast artifacts in CT-based

attenuation correction of PET images using an automated segmentation

algorithm. European Journal of Nuclear Medicine & Molecular Imaging,

2008. 35(10): p. 1812-1823.

106. Middleton, I. and R.I. Damper, Segmentation of magnetic resonance images

using a combination of neural networks and active contour models. Medical

Engineering & Physics, 2004. 26(1): p. 71-86.

References

170

107. Ramme, A., et al., Semi-automated Phalanx Bone Segmentation Using the

Expectation Maximization Algorithm. Journal of Digital Imaging, 2009.

22(5): p. 483-491.

108. Ramme, A., et al., EM Segmentation of the Distal Femur and Proximal Tibia:

A High-Throughput Approach to Anatomic Surface Generation. Annals of


109. Chen, T., L. Serra, and H. Ng, Surface extraction: dividing voxels.

International Congress Series, 2004. 1268: p. 225-230.

110. Newman, T.S. and H. Yi, A survey of the marching cubes algorithm.

Computers & Graphics, 2006. 30(5): p. 854-879.

111. Lucas, L., D. Gillard, and Y. Rémion, A new unsupervised cube-based

algorithm for iso-surface generation. Computer Networks and ISDN

Systems, 1997. 29(14): p. 1737-1744.

112. Lorensen, W.E. and H.E. Cline, Marching cubes: A high resolution 3D

surface construction algorithm. SIGGRAPH Comput. Graph., 1987. 21(4): p.

163-169.

113. Popescu, F., et al., A new method to compare planned and achieved position

of an orthopaedic implant. Computer Methods and Programs in Biomedicine,

2003. 71(2): p. 117-127.

114. Tocheri, M.W., et al., A 3D quantitative comparison of trapezium and

trapezoid relative articular and nonarticular surface areas in modern

humans and great apes. Journal of Human Evolution, 2005. 49(5): p. 570-

586.

115. Li, K., et al., Mapping ligament insertion sites onto bone surfaces in knee by

co-registration of CT and digitization data. Journal of Biomechanics, 2009.

42(15): p. 2624-2626.

References

171

116. Schmutz, B., K.J. Reynolds, and J.P. Slavotinek, Development and validation

of a generic 3D model of the distal femur. Computer Methods in

Biomechanics & Biomedical Engineering, 2006. 9(5): p. 305-312.

117. Ruedi, T.P. and W.M. Murphy, AO Principles of Fracture Management.

Thieme Stuttgart, New York, 2001: p. 50-51.

118. Haraguchi, K., et al., Comparison of Fit and Fill Between Anatomic Stem and

Straight Tapered Stem Using Virtual Implantation on the ORTHODOC

Workstation. Computer Aided Surgery, 2001. 6(5): p. 290-296.

119. Genant, H. and S. Majumdar, High-resolution magnetic resonance imaging of

trabecular bone structure. Osteoporosis International, 1997. 7(0): p. 135-139.

120. Majumdar, S., et al., Magnetic resonance imaging of trabecular bone

structure in the distal radius: Relationship with X-ray tomographic

microscopy and biomechanics. Osteoporosis International, 1996. 6(5): p. 376-

385.

121. Wehrli, F.W., et al., Role of Magnetic Resonance for Assessing Structure and

Function of Trabecular Bone. Topics in Magnetic Resonance Imaging, 2002.

13(5): p. 335-355.

122. Liu, J., et al., Rigid model-based 3D segmentation of the bones of joints in

MR and CT images for motion analysis. Medical Physics, 2008. 35(8): p.

3637-3649.

123. Pillai, R.R., et al., MRI-based modeling for evaluation of in vivo contact

mechanics in the human wrist during active light grasp. Journal of

Biomechanics, 2007. 40(12): p. 2781-2787.

124. Moro-oka, T.-a., et al., Can magnetic resonance imaging-derived bone

models be used for accurate motion measurement with single-plane three-

dimensional shape registration? Journal of Orthopaedic Research, 2007.

25(7): p. 867-872.

References

172

125. DeFrate, L.E., et al., In vivo tibiofemoral contact analysis using 3D MRI-

based knee models. Journal of Biomechanics, 2004. 37(10): p. 1499-1504.

126. Chen, B., et al., Combined MR imaging towards subject-specific knee contact

analysis. Visual Computer, 2011. 27(2): p. 121-128.

127. Hao, Z., et al., A finite element 3D model of in vivo human knee joint based

on MRI for the tibiofemoral joint contact analysis. 2007. p. 616-622.

128. Gold, G.E., et al., Musculoskeletal MRI at 3.0 T: Initial Clinical Experience.

American Journal of Roentgenology, 2004. 183(5): p. 1479-1486.

129. Tsurusaki, M., et al., 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): p. 314-320.

130. Ramalho, M., et al., Liver MR Imaging: 1.5T versus 3T. Magnetic Resonance

Imaging Clinics of North America, 2007. 15(3): p. 321-347.

131. Rakow-Penner, R., et al., Relaxation times of breast tissue at 1.5T and 3T

measured using IDEAL. Journal of Magnetic Resonance Imaging, 2006.

23(1): p. 87-91.

132. Frayne, R., et al., Magnetic resonance imaging at 3.0 Tesla: Challenges and

advantages in clinical neurological imaging. Investigative Radiology, 2003.

38(7): p. 385-402.

133. Beyersdorff, D., et al., 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): p. 1214-1220.

134. Schindera, S.T., et al., Abdominal Magnetic Resonance Imaging at 3.0 T:

What Is the Ultimate Gain in Signal-to-Noise Ratio? Academic Radiology,

2006. 13(10): p. 1236-1243.

135. Bowers, M.E., et al., Effects of ACL interference screws on articular

cartilage volume and thickness measurements with 1.5 T and 3 T MRI.

Osteoarthritis and Cartilage, 2008. 16(5): p. 572-578.

References

173

136. Chen, C.A., et al., Cartilage morphology at 3.0T: Assessment of three-

dimensional magnetic resonance imaging techniques. Journal of Magnetic

Resonance Imaging, 2010. 32(1): p. 173-183.

137. Stehling, C., et al., Three-Tesla Magnetic Resonance Imaging of the Wrist:

Diagnostic Performance Compared to 1.5-T. Journal of Computer Assisted

Tomography, 2009. 33(6): p. 934-939.

138. Lambert, A., et al., Rotator cuff tears: value of 3.0T MRI. Journal of

Radiology, 2009. 90: p. 583-588.

139. Duewell, S.H., et al., Musculoskeletal MR imaging at 4 T and at 1.5 T:

comparison of relaxation times and image contrast. Radiology, 1995. 196(2):

p. 551-555.

140. Bernstein, M.A., J. Huston, and H.A. Ward, Imaging artifacts at 3.0T.

Journal of Magnetic Resonance Imaging, 2006. 24(4): p. 735-746.

141. Kuo, R., et al., 3.0 Tesla imaging of the musculoskeletal system. Journal of

Magnetic Resonance Imaging, 2007. 25(2): p. 245-261.

142. Soher, B.J., B.M. Dale, and E.M. Merkle, A Review of MR Physics: 3T versus

1.5T. Magnetic Resonance Imaging Clinics of North America, 2007. 15(3): p.

277-290.

143. Ramnath, R.R., 3T MR imaging of the musculoskeletal system (Part II):

clinical applications. Magnetic Resonance Imaging Clinics of North

America, 2006. 14(1): p. 41-62.

144. Lawrence, N.T., 3T MRI in clinical practice. Applied Radiology, 2005. 34(1):

p. 8.

145. Mosher, T.J., Musculoskeletal Imaging at 3T: Current Techniques and Future

Applications. Magnetic Resonance Imaging Clinics of North America, 2006.

14(1): p. 63-76.

146. Norris, D.G., High field human imaging. Journal of Magnetic Resonance

Imaging, 2003. 18(5): p. 519-529.

References

174

147. Gold, G.E., et al., Musculoskeletal MRI at 3.0 T: Relaxation Times and Image

Contrast. American Journal of Roentgenology, 2004. 183(2): p. 343-351.

148. Stanisz, G.J., et al., T1, T2 relaxation and magnetization transfer in tissue at

3T. Magnetic Resonance in Medicine, 2005. 54(3): p. 507-512.

149. Delfaut, E.M., et al., Fat Suppression in MR Imaging: Techniques and

Pitfalls. Radiographics, 1999. 19(2): p. 373-382.

150. Abduljalil, A.M. and P.-M.L. Robitaille, Macroscopic Susceptibility in Ultra

High Field MRI. Journal of Computer Assisted Tomography, 1999. 23(6): p.

832-841.

151. Peh, W.C.G. and J.H.M. Chan, Artifacts in musculoskeletal magnetic

resonance imaging: identification and correction. Skeletal Radiology, 2001.

30(4): p. 179-191.

152. Firbank, M.J. and et al., A comparison of two methods for measuring the

signal to noise ratio on MR images. Physics in Medicine and Biology, 1999.

44(12): p. N261.

153. Dietrich, O., et al., 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): p. 375-385.

154. Henkelman, R.M., Measurement of signal intensities in the presence of noise

in MR images. Medical Physics, 1985. 12(2): p. 232-233.

155. Edelstein, W.A., P.A. Bottomley, and L.M. Pfeifer, A signal-to-noise

calibration procedure for NMR imaging systems. Medical Physics, 1984.

11(2): p. 180-185.

156. Stadler, A., et al., Artifacts in body MR imaging: their appearance and how

to eliminate them. European Radiology, 2007. 17(5): p. 1242-1255.

157. Odille, F., et al., Generalized MRI reconstruction including elastic

physiological motion and coil sensitivity encoding. Magnetic Resonance in

Medicine, 2008. 59(6): p. 1401-1411.

References

175

158. Besl, P.J. and N.D. McKay, A Method for Registration of 3-D Shapes. IEEE

Transaction on Pattern Analysis and Machine Intelligence, 1992. 14(2): p.

239-256.

3D Reconstruction of Long Bones Utilising Magnetic ... · Abstract IV usability of MRI as a...

Documents

Transcript of 3D Reconstruction of Long Bones Utilising Magnetic ... · Abstract IV usability of MRI as a...