Thesis

Automated Assessment of
Acetabular Wear From

Anteroposterior Radiographs
using Active Ellipses
Charles Stuart Kerrigan
A Thesis presented for the degree of

Doctor of Philosophy
Computer Vision and Imaging Group

Department of Applied Computing
University of Dundee
Dundee
Scotland
September 8, 2005
Contents
1 Introduction
1.1
Motivation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Aims and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3
Contents and contributions . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4
Structure of this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Background
2.1
Total Hip Replacements . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2
Failure of THRs
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1
Radiolucency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.2
Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.3
Acetabular Wear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3
Radiography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3 Related Work
22
3.1
2D Wear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2
3D Wear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3
2D Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4
3D Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4 Modelling
4.1
40
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
ii
4.2
Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3
Shape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3.1
Fitting Ellipses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3.2
Modelling Ellipses on Radiographs . . . . . . . . . . . . . . . . . . . 46
4.4
Appearance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.5
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5 Parameter Estimation
5.1
53
Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.1.1
Template Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.1.2
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2
Data Points and Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.3
Robust Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.4
Review of Ellipse Error Functions . . . . . . . . . . . . . . . . . . . . . . . . 63
5.5
5.6
5.7
5.4.1
Algebraic Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.4.2
Weighted Algebraic Distance . . . . . . . . . . . . . . . . . . . . . . 64
5.4.3
Angular Bisector of the Foci . . . . . . . . . . . . . . . . . . . . . . . 64
5.4.4
Orthogonal Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.5.1
Synthetic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.5.2
Radiographic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.6.1
Synthetic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.6.2
Radiographic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6 Eccentricity Error and Pose Estimation
84
6.1
Eccentricity Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2
Pose Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3
6.2.1
CAD Model Based Pose Estimation . . . . . . . . . . . . . . . . . . 88
6.2.2
Pose of a Circle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
iii
6.4
Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7 Measuring Acetabular Wear
98
7.1
Experimental Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2
Uncorrected Wear Experiment . . . . . . . . . . . . . . . . . . . . . . . . . 99

7.2.1
Active Ellipses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.2.2
Comparative Method
7.2.3
Measuring Wear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.2.4
Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.2.5
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
. . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3
Normals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.4
Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.5
Corrected Wear Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.6
7.5.1
Correction Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.5.2
Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.5.3
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
8 Conclusions
122
A The Ellipse
126
A.1 Explicit Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

A.2 Implicit Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
A.3 Matrix Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
A.4 Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
B System Overview
131
B.1 Overall System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

B.2 Training Subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
B.2.1 Annotation Pseudocode . . . . . . . . . . . . . . . . . . . . . . . . . 133
B.2.2 Femoral Head Annotation Pseudocode . . . . . . . . . . . . . . . . . 133
B.2.3 Acetabular Rim Annotation Pseudocode . . . . . . . . . . . . . . . . 133
B.2.4 Model Calculation Code . . . . . . . . . . . . . . . . . . . . . . . . . 134
iv
B.3 Contour Localization Subsystem . . . . . . . . . . . . . . . . . . . . . . . . 135

B.4 Wear Measurement Subsystem . . . . . . . . . . . . . . . . . . . . . . . . . 138
C Publications resulting from this work
139
List of Figures
1.1
A total hip replacement. (a) The femoral head placed on the neck of the
stem component. (b) The acetabular component, also called acetabular
cup, usually made from polyethylene. Of particular note is the circular
metallic wire rim. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
The femoral head and acetabular cup as they manifest on a radiograph
1.3
Two synthetic views simulating an extreme amount wear over time. (a)
. .
centre of femoral head close to centre of acetabular rim and (b) obvious
indication that the femoral head has penetrated the acetabular cup . . . . .
2.1
Region of interest of a healthy hip. 1) Acetabulum, the large cup-shaped

structure on the lateral surface of the pelvis. 2) Femoral head which articulates within the acetabulum . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2
A total hip replacement. 1) Acetabular component, also called acetabular

cup, usually made from polyethylene. Of particular note is the elliptical
metallic wire rim. 2) Metal femoral head which articulates within the acetabular cup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3
Region of interest of a total hip replacement. 1) Acetabular rim wire that

marks the opening of the acetabular cup. 2) Edge marker used by Charnley
to measure acetabular wear. 3) Femoral head, the spherical component of
the total hip replacement. 4) Neck of the prosthesis. 5)Stem of the prosthesis
2.4
Diagram of a cemented THR . . . . . . . . . . . . . . . . . . . . . . . . . . 10
vi
2.5
Components of a cementless THR include (a) the acetabular cup consisting

of an outer shell with porous coating and screw to allow fixation in bone and
a polyethylene liner (b) porous coating of the stem of the femoral component
(c) bonding of the porous coating to the bone . . . . . . . . . . . . . . . . . 11
2.6
(a) Seven Gruen zones around the femoral component (b) Three DeLee
zones around the acetabular component. Image taken from Gruen et al.[41]. 13
2.7
Visualisation of the modes of failure. Image taken from Gruen et al.[41]. . . 14
2.8
A fractured stem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.9
Radiolucency of the femoral stem . . . . . . . . . . . . . . . . . . . . . . . . 15
2.10 Obtaining a THR radiograph through a simplified x-ray imaging system . . 19

2.11 The symphysis pubis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1
The narrowest and widest areas in the weight-bearing area of the prosthesis,
also known as the thinness and thickness . . . . . . . . . . . . . . . . . . . . 24
3.2
The ischial tuberosities and a line through the caudal points are used to
determine the angle of wear. . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3
Medial (m), lateral (l) and central points (c) with the initial estimate of the
centre of the femoral head determined by a circle fit . . . . . . . . . . . . . 27
3.4
Circle fitting in Polyware . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5
Model created using Polyware showing the AP (left) and ML (right) views . 32
3.6
Migration measured by Sutherland et al.s method. m1 is the caudal most

point of the adjacent teardrop. m2 is the ischium on the Kohler line. m3 is
the ilium on the Kohler line. . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.7
Migration measured by Nunns method. T) A line passing through the

caudal points of the teardrop to measure y-distance. P) Orthogonal to Line
T to measure x-distance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.8
EBRA method with reference lines placed on pelvis and annotated points
on both the femoral head and acetabular rim. . . . . . . . . . . . . . . . . . 34
3.9
An example radiograph with tantalum markers shown . . . . . . . . . . . . 36
3.10 Insertion of tantalum markers into (a) the femur (b) the acetabulum . . . . 37
vii
3.11 (a) Typical setup for obtaining an RSA radiograph. (b) Hip replacement
positioned at the intersection of both x-ray sources . . . . . . . . . . . . . . 37
4.1
Comparison of two fits to the same points using LSO initialised from a circle
fit and a standard LSA fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2
(a) Set of data points on the femoral head. (b) An LSO fit with 30 normals
extended from the contour (10 more were discounted) . . . . . . . . . . . . 47
4.3
(a) Set of data points on the acetabular rim. (b) An LSO fit with 40 normals
extended from the contour . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4
Plot of the mean profiles, with error bars showing 1 SD, obtained from 50
postoperative radiographs for (a) the femoral head and (b) the acetabular
rim, where k1 = k2 = 31 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.5
Location of datapoints on (a) the femoral head using the mean and covariance of the femoral head profiles, where m1 = 111 and (b) the acetabular
rim using the mean and covariance of the acetabular rim profiles where
m2 = 151. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.6
The Active Ellipse Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1
Template image and subsampled template . . . . . . . . . . . . . . . . . . . 55
5.2
Successful initialization by template and resultant convergence . . . . . . . 55
5.3
Region of interest and full view of radiograph where failure to localise

femoral head contour is due to quantisation effect caused during scanning.
This manifests itself as a streaking effect across the entire radiograph. . . . 56
5.4
(a) Region of interest of two radiographs with noise shown after histogram
equalisation and (b) quantisation effect caused during scanning. This prevents localization of the contours. . . . . . . . . . . . . . . . . . . . . . . . . 56
5.5
Region of interest of two radiographs with poor levels of contrast. (a)

This radiograph was not scanned on a backlit scanner, resulting in the dark
image and (b) the femoral head is almost occluded by the surrounding bone.
The template is unable to locate the femoral head and instead favours the
symphisis pubis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
viii
5.6
Region of interest of two radiographs with poor levels of contrast. (a) Poor
contrast prevents accurate localisation of the femoral head. (b) a poorly
contrasted radiograph containing a metal backed acetabular cup. . . . . . . 57
5.7
(a) Set of data points on the femoral head. (b) The resultant LS fit . . . . . 59
5.8
A single outlying point distorts the LS fit (dashed line) from the original
ellipse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.9
The LMedS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.10 Location of the closest point of intersection of the angular bisector (xf ) . . 65
5.11 Initialisation (x0 ) and successful location of the orthogonal contacting point
(x) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.12 The two ellipses used to generate synthetic data. Both were centred at
the origin and aligned with the image axes. (a) The less eccentric ellipse,
a = 333, b = 250. (b) The more eccentric ellipse, a = 300, b = 60.
. . . . . 69
5.13 Data points created from the more eccentric ellipse. (a) Gaussian noise
( = 10). (b) Half low variance Gaussian noise ( = 5) and half high variance Gaussian noise ( = 20 in this visualisation). (c) Structured outliers
sampled from noisy line segments . . . . . . . . . . . . . . . . . . . . . . . . 70
5.14 An acetabular rim with eccentricity 1.0. The method cannot localise such
a contour. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.15 Centre errors for (a) less eccentric and (b) more eccentric synthetic ellipse
data as a function of . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
data with Gaussian outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
data with structured outliers . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.18 Standard deviation of centre errors for (a) less eccentric and (b) more eccentric synthetic ellipse data as a function of . . . . . . . . . . . . . . . . 75
5.19 Standard deviation of centre errors for (a) less eccentric and (b) more eccentric synthetic ellipse data with Gaussian outliers . . . . . . . . . . . . . . 76
5.20 Standard deviation of centre errors for (a) less eccentric and (b) more eccentric synthetic ellipse data with structured outliers . . . . . . . . . . . . . 77
ix
data with Gaussian outliers, visualising only foci bisector and orthogonal
distance error functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
data with structured outliers, visualising only foci bisector and orthogonal
distance error functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.23 Failure of the method as the highly eccentric rim is at such an extreme
orientation and thus too far from the mean rim parameters . . . . . . . . . 80
5.24 Failure of the method as the upper end point of the ellipse is occluded by
clutter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.25 Failure of the method as the fitting has favoured the near circular wire
marker on the periphery of the acetabular cup instead of the rim . . . . . . 81
5.26 Failure of the method as the fitting has favoured the near circular wire
marker on the periphery of the acetabular cup instead of the rim, as the
rim wire marker is occluded . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.27 An eccentric rim with (a) failed algebraic, (b) successful foci bisector and
(c) successful geometric fits. . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.1
Projecting both a circle and an ellipse from a circle . . . . . . . . . . . . . . 84
6.2
The acetabular cup with (a) circular rim and (b) rotated in depth to produce
an elliptical rim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.3
Side view of the object and image plane, aligned with y 0 to show the projection of the entire minor axis. This shows the projection of the minor axis b,
with eccentricity error y0 shown as the distance between projected ellipse
centre B and actual projection of the centre C. In this instance =. . . . 86
6.4
Visualisation in 3-D of x-ray system from focus to object in a case where the
ellipse is not at the beam centre. Finally 6= for visualisation purposes.
When = this diagram collapses down. . . . . . . . . . . . . . . . . . . . 87
6.5
Typical geometry for recovering the pose of a circle from a camera setup.
Note x0 and y 0 are aligned with the axes of the ellipse on the film.
. . . . . 89
6.6
(a) A rotated and translated ellipse and the resultant ellipse axes x0 , y 0 ,
z 0 (z 0 not pictured) obtained from using the eigenvalues of the symmetric
matrix as coefficients as outlined in Equation 6.7. When expressed in the
x0 , y 0 , z 0 coordinate system the ellipse is aligned with the axes and centred
on the z 0 axis (c) The resultant axes x00 , y 00 , z 00 (z 00 not pictured) obtained
by rotating in depth using in Equation 6.8 . . . . . . . . . . . . . . . . . 90
6.7
Correction of centres where (a) FFD=1300 and OFD=180 (b) FFD=1000

and OFD=225 (c) FFD=900 and OFD=270 . . . . . . . . . . . . . . . . . . 93
6.8
Correction of centres where (a) FFD=1300 and OFD=180 (b) FFD=1000

and OFD=225 (c) FFD=900 and OFD=270 . . . . . . . . . . . . . . . . . . 94
6.9
Distribution of rotations in depth obtained for (a) Year 1 and (b) Year 5
radiographs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.10 Distribution of the difference between rotations in depth obtained for Year
1 and Year 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.11 Radiograph with (a) lowest rotation in depth (53.2 ) and (b) highest rotation in depth (85.9 ) from dataset . . . . . . . . . . . . . . . . . . . . . . . 96
7.1
Average parameters visualised on a radiograph for (a) the femoral head and
(b) the acetabular rim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.2
(a) A set of data points on the femoral head. (b) A set of data points on
the acetabular rim. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3
(a) Vector obtained at Year 1 (b) Vector obtained at Year 5 (c) Wear
computed between the two vectors . . . . . . . . . . . . . . . . . . . . . . . 104
7.4
Bland and Altman plots for Annotator 1 and the Active Ellipse Method . . 106
7.5
7.6
7.7
7.8
Distribution of wear over 4 years for the active ellipses method . . . . . . . 108
7.9
Distribution of wear over 4 years for annotator 1 . . . . . . . . . . . . . . . 109
7.10 Distribution of wear over 4 years for annotator 2 . . . . . . . . . . . . . . . 109

7.11 Distribution of wear over 4 years for annotator 3 . . . . . . . . . . . . . . . 110
xi
7.12 Distribution of wear over 4 years for annotator 4. Outlying measurements

are not shown here . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.13 Distribution of wear over 4 years for annotator 4 showing outlying measurements not shown in Figure 7.12 . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.14 An incorrect measurement performed on a year 5 radiograph by annotator 4 114
7.15 Plot of normals extended from an active ellipse against repeatability of
paired measurements of acetabular wear . . . . . . . . . . . . . . . . . . . . 116
7.16 Distribution of differences between uncorrected and corrected wear values . 119
A.1 Projecting both a circle and an ellipse from a circle . . . . . . . . . . . . . . 126
A.2 Visualisation of the 5 ellipse parameters . . . . . . . . . . . . . . . . . . . . 127
A.3 Canonical example of an ellipse, where xc = yc = = 0. . . . . . . . . . . . 127
B.1 Data flow diagram of the training subsystem . . . . . . . . . . . . . . . . . 132
B.2 Data flow diagram of the Contour Localization subsystem and subsequent
eccentricity error correction . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
B.3 Data flow diagram of the wear calculation subsystem and subsequent eccentricity error correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
xii
List of Tables
1
List of nomenclature included in this thesis . . . . . . . . . . . . . . . . . . xix
Definitions for medical terms used in this thesis. . . . . . . . . . . . . . . . xxii
2.1
Modes of Cemented Femoral Stem Loosening . . . . . . . . . . . . . . . . . 14
5.1
Template Matching Success . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.1
Eccentricity error values for given parameters. . . . . . . . . . . . . . . . . . 87
7.1
Variables for the active ellipses used in obtaining wear measurements . . . . 101
7.2
Limits of agreement of each annotator with the active ellipses method, mean
limits of agreement and alpha trimmed mean where = 0.5 . . . . . . . . . 110
7.3
Mean and standard deviation of differences between paired wear measurements with no eccentricity correction . . . . . . . . . . . . . . . . . . . . . . 111
7.4
Mean and standard deviation of centroid displacements in Year 1 radiographs with no eccentricity correction . . . . . . . . . . . . . . . . . . . . . 112
7.5
Mean and standard deviation of centroid displacements in Year 5 radiographs with no eccentricity correction . . . . . . . . . . . . . . . . . . . . . 113
7.6
Observed Parameters of 2-D Wear . . . . . . . . . . . . . . . . . . . . . . . 113
7.7
Mean and standard deviation of uni-radiographic measurements performed

on differing resolutions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.8
Mean and standard deviation of uni-radiographic measurements performed

on differing resolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.9
Variables for the active ellipses used in the resolution experiment at 300 dpi 117
xiii
7.10 Mean and standard deviation of differences between paired active ellipse
wear measurements with varying eccentricity correction . . . . . . . . . . . 119
7.11 Observed parameters of 2D Wear with eccentricity correction . . . . . . . . 119
7.12 Mean and standard deviation of difference between uncorrected and corrected wear measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
xiv
Acknowledgements
I would like to acknowledge the support of my supervisors, Stephen McKenna, Ian Ricketts
and Carlos Wigderowitz. My thanks go to my mother and father, Ellen and Charles,
who have supported me unwaveringly through this journey. I am also grateful to Steven,
Gregor, Mary and all the others who have helped me keep a sense of perspective during the
last 4 years. Finally I would like to thank the staff and students of both the Departments
of Applied Computing and Orthopaedics and Trauma Surgery and in particular mention
the following individuals who contributed to this project:
Lynda Cochrane for her aid with the statistics contained in Chapter 7
Karthik Pingle, Bharti Rajput, Karen Duffy and Graeme Fenwick who were patient
enough to perform over 200 annotations of radiographs
Tony Mochan and Jeremy Martindale for allowing me to obtain the datasets of
radiographs
Paul Rosin and Sung Ahn for their correspondance and advice on implementing the
ellipse fitting algorithms
xv
Declaration
I hereby declare that the work described in this thesis is my own; that I am the author of
this thesis; that it has not previously been put forward in submission for any other degree
or qualification; and that I have consulted the references listed herein.
Signed
Charles Stuart Kerrigan
xvi
Supervisors Declaration
We certify that Mr Charles Stuart Kerrigan has satisfied all the terms and conditions of
the regulations made under Ordinances 12 and 39 and has completed the required nine
terms of research to qualify in submitting this thesis in application for the degree of Doctor
of Philosophy.
Signed
Stephen McKenna, Ian Ricketts, Carlos Wigderowitz
xvii
Nomenclature
This is a list of terminology used in this thesis, along side a brief explanation.
Parameter
Description
Volume of wear
Femoral head radius
Linear wear (mm)
Angle between acetabular rim major axis and linear wear tract
Standard Deviation
A, B, C, D, E, F
Conic coefficients
Vector containing the 6 ellipse coefficients
Design vector containing [x2 , xy, y 2 , x, y, 1]
Covariance matrix of normalized derivative profiles
First order grey level derivative
First order normalized grey level derivative
Mean normalized first order grey level derivative
Ti,j
Template pixel accessed via indices i, j
Ix,y
Image pixel x, y
k1 , k2
Length of femoral head and acetabular rim model normals
m1 , m2
Length of femoral head and acetabular rim search normals
N1 , N 2
No. of femoral head and acetabular rim ellipse normals
Number of images in training set
Number of points for fitting ellipse
xc , yc
Centre coordinate parameters of an ellipse
a, b
Major and minor sub-axes parameters of an ellipse

continued on next page
xviii
continued from previous page

Parameter
Description
Rotation parameter of an ellipse
Eccentricity of ellipse
f1 , f2
Vector of coordinates corresponding to the ellipse foci
a
, b,
Mean ellipse parameters
a, b, xc , yc ,
Convergence criteria
Number of minimal subsamples for LMedS fitting
Step size for minimization during Least Squares Orthogonal fit
Rotation in depth of a circle
Jacobian matrix for minimizing

orthogonal contacting points conditions
x0 , y 0 , z 0
Coordinate system aligned with the axes of target ellipse
x , y , z
Coordinate system aligned with with circular object
x0 , y0
Eccentricity error along the x0 and y 0 axes
Camera rotation about the x axis (assumed to equal )
Symmetric matrix of ellipse coefficients
e1 , e2 , e3
Eigenvectors obtained from symmetric matrix
1 , 2 , 3
Eigenvalues of symmetric matrix
R1
Rotation and translation matrix to transform from x, y, z to x0 , y 0 , z 0
R2
Rotation matrix constructed using
Normal to the circle
Table 1: List of nomenclature included in this thesis
xix
Glossary
This is a list of medical terminology in this thesis with an explanatory description.
Term
Definition
Acetabular Cup
Polyethylene cup shaped component designed

to replace the acetabulum in a THR.
Acetabular Rim
Wire placed on the mouth or rim of the cup

for reference on radiographs.
Acetabular Edge Wire
Wire placed on the periphery of the cup.

for reference on radiographs.
Acetabulum
A large cup-shaped bone at the hip joint

which the femoral head articulates with.
Anteroposterior
Front-to-back direction.
Aseptic Loosening
Loosening of a THR that is not caused by infection.
Caudal
Downwards direction, towards the tail or bottom end of patient.
Cementless
THR that does not require bone cement, using

tightness of fit and coating instead to aid fixation.
Cemented
THR cushioned by a layer of cement.
Coronal Plane
Plane that passes through the body dividing

the body into front and back.
Cranial
Upwards direction, towards the skull or cranium.
DeLee Zones
Delineated zones for the assessment of radiolucency

around the acetabular component
Femoral Head
Refers to
a) in original hip joint the rounded head
xx

Term
Definition
that interacts with the acetabulum
b) the spherical component of a
THR that articulates with the acetabular cup.
Femur
Tubular bone on the upper leg which articulates

with the hip at the acetabulum.
Film Focus Distance (FFD)
Distance from the x-ray focus to the film
Gruen Zones
Delineated zones for the assessment of radiolucency

around the femoral component
Hybrid
THR where one component is cementless and the other

is cemented.
Ischial tuberosities
Bones on the buttocks used as reference points.
Linear Wear
Linear wear is the maximum depth of the largest pit

on the acetabular cup surface.
Mediolateral
Right-to left or left-to right direction.
Migration
Change in the position of the femoral or

acetabular components.
Novel
The article in question has not been seen before,

for example novel images have not been
used to train the system in this thesis.
Object to Film Distance (OFD)
Distance from object reference

point to the film
Osteolysis
Condition that causes bone resorption to exceed

bone formation and thus a loss of bone volume
Primary THR
Initial THR received by patient.
Radiograph
The resultant radiographic image on x-ray film

showing the structures of the imaged object.
Radiolucency
Dark area on radiograph indicating lack of

attenuation of x-rays (e.g. cavity)
Revision THR
THR that replaces either a failed primary or revision THR.
Symphysis pubis
Joint formed by the bodies of the pubic bones.

xxi

Term
Definition
Beam centre of pelvic x-rays are centred
on the symphysis pubis.
Total Hip Replacement (THR)
Replacement of existing hip joint

with a prosthetic hip joint.
Volumetric Wear
Estimation of the volume of polyethylene particles

worn away.
Wear
Wear of the THR through adhesion, abrasion and fatigue

produces submicron particles, known as wear particles.
Polyethylene from the acetabular cup is the major source
of wear particles.
Table 2: Definitions for medical terms used in this thesis.
xxii
Abstract
Acetabular wear is the major cause of aseptic loosening of total hip replacements, leading
to costly revision surgery and patient discomfort. To measure acetabular wear from radiographs manually is a labour-intensive task and most of the current computerised methods
either do little to alleviate this problem or require costly equipment that prevents use in
standard clinical practice.
The development of a new and completely automated method for measuring acetabular wear from anteroposterior radiographs is described. This method uses the standard
PC and x-ray acquisition setups, ensuring it remains feasible for the clinical setting. Parametric active shape models, active ellipses, are used to automatically locate two elliptical
contours in a given radiograph. By comparing the position of these contours on two radiographs taken at different periods in the prosthesis lifespan it is possible to compute a
rate of wear. The problems of modelling both the shape and appearance of the contours
are addressed. Template matching is used to provide an initial estimate as to the location
of the contours. The robust statistical Least Median of Squares method is used to localize
the contours.
This method is then applied to performing measurements on a clinical dataset, with an
evaluation of the proposed methods repeatability. A labour-intensive method involving
manual annotation of digital radiographs is compared to the novel method. The latter
is found to be more repeatable. Additionally, the eccentricity error of using the highly
eccentric elliptical acetabular rim marker is investigated. A method of correcting for this
error is suggested and the impact on uncorrected measurements is examined.
Chapter 1
Introduction
Total Hip Replacement (THR) surgery is an increasingly common operation whereby a
damaged hip is replaced by a prosthetic one. The spherical femoral head articulates
within the polyethylene cup, both shown in Figure 1.1, to simulate the working of a
normal hip joint. Figure 1.2 shows an example of how the components manifest on a
typical anteroposterior (AP) radiograph, which is a radiograph where the x-ray machine
is in front of the patient and the x-rays pass through to the back.
The THR operation stems from the works of Sir John Charnley and has paved the way
for total shoulder and knee replacement surgery. In the 1950s and early 1960s while based
at Wrightington Hospital, Charnley sought a method of replacing both the femoral head
and acetabulum of the hip. He first attempted to replace the acetabular socket with a
Teflon implant, but ultimately moved on to polyethylene fixed with polymethylmethacrylate, a substance used by dentists and known as bone cement. By 1961, Charnleys operations were proving successful and to this day his techniques and designs are still used.
However numerous variant designs of prosthesis are on the market now and the exact
model of prosthesis used often depends on patient and surgeon preference.
1.1
Motivation
Approximately 12.5% of THR operations are revision hip replacements, where the original
hip replacement has failed. Using conventional methods to assess THRs through radiographs the data on poorly performing models can only be found after an extended period
(a)
(b)
Figure 1.1: A total hip replacement. (a) The femoral head placed on the neck of the stem
component. (b) The acetabular component, also called acetabular cup, usually made from
polyethylene. Of particular note is the circular metallic wire rim.
Figure 1.2: The femoral head and acetabular cup as they manifest on a radiograph
of time during which a poorly performing model can have been used in many operations.
The objective of this work is to create a new system for assessing hip implants that
should be as automated as possible without using specialist equipment that would render
it unfeasible in a standard clinical setting. The system should be more repeatable than
more labour-intensive alternatives. A system with sufficient accuracy to provide early
measurements of acetabular wear would be a valuable tool to evaluate prosthesis designs
and has clear economic benefits, as well as increasing patient confidence and comfort.
Wear manifests itself on radiographs as the penetration of the femoral head into the
acetabular cup. An example of this can be seen in Figure 1.3, where when comparing the
first and second images it becomes clear the femoral head has penetrated the acetabular
rim. A more detailed explanation of loosening, migration and acetabular wear is presented
in Chapter 2, but this thesis focuses on assessing acetabular wear as it is responsible for the
majority of failures. It also investigates a potential source of systematic error in measuring
acetabular wear in order to ensure that measuring wear remains feasible.
(a)
(b)
Figure 1.3: Two synthetic views simulating an extreme amount wear over time. (a) centre
of femoral head close to centre of acetabular rim and (b) obvious indication that the
femoral head has penetrated the acetabular cup
1.2
Aims and Objectives
1.3
Contents and contributions
This thesis details the development of a computer vision-based system for the automated
measurement of acetabular wear in AP radiographs. The system is capable of providing
measurements without user-interaction, or failing that minimal user-interaction. Previous
systems are either labour-intensive or so costly that their use in a clinical setting is often
not feasible. The development of a new system requires consideration of several important
issues which are influenced by the requirements of the system. These issues include modelling, segmentation and consideration of any systematic and random errors introduced
into the measurements. Modelling examines how to represent the shape of the femoral
head and acetabular rim and the appearance of the contours. Segmentation considers how
to allow the correct solution to be found in the presence of false matches to the model.
Finally one must ascertain whether error inherent in the method is small enough to allow
measurements with the desired level of accuracy.
This thesis makes the following novel contributions: Statistical models of the profiles of the femoral head and acetabular rim contours
Robust detection of femoral head and acetabular rim elliptical contours
Evaluation of Least Squares (LS) combined with iterative geometric ellipse fitting
Evaluation of Least Median of Squares (LMedS) ellipse fitting using a geometric
distance error function
Estimation and correction of eccentricity errors for the acetabular rim contour centre
Pose estimation of the acetabular rim
Evaluation of method when used on clinical data
Effect of resolution on standard deviation of displacement measurements
1.4
Structure of this Thesis
The content of this thesis is organised into 8 chapters. Chapter 2 contains a brief review
of THR surgery and the chief causes of failure, including expanding on the concept of
acetabular wear. Existing work relating to automated assessment of total hip replacements
is examined in Chapter 3. The issues concerned with modelling the shape and appearance
of the contours of the relevant structures in the radiographs are considered in Chapter
4, whilst the issues with and solutions to applying the model to a novel set of data are
outlined in Chapter 5. Chapter 6 considers an often overlooked source of systematic error,
known as eccentricity error, that is introduced into the method. It outlines a method to
correct for this error. The performance of the system measuring acetabular wear in clinical
data is outlined in Chapter 7 by an assessment of the systems repeatability. Corrected
measurements are compared with uncorrected ones to obtain an estimate of the systematic
errors magnitude. Finally the conclusions and suggestions for future expansion of this
work are presented in Chapter 8.
Appendix A gives additional background on the ellipse, giving explicit and implicit
forms of representing the ellipse (both of which are used in this thesis). Appendix B
on the other hand gives a high-level software engineers view of the method described in
this thesis. Dataflow diagrams and pseudocode are provided. This appendix should help
clarify the workings of the entire system.
Chapter 2
Background
This chapter describes the purpose and process of total hip replacement (THR) and the
causes of failure. It then explains the need for radiographic assessment in determining
prosthesis failure. It has been asserted by Bould et al. [11] that radiographic assessment
depends on both the reproducibility of radiographic images and the accuracy or reproducibility of radiographic measurement. Therefore it is not just methods for performing
measurements that must be examined, but also the process of acquiring radiographic images of a total hip replacement. This chapter highlights variables that might affect the
reproducibility of radiographs throughout the lifespan of the prosthesis.
2.1
Total Hip Replacements
Over 40,000 THRs [51] are performed in the United Kingdom annually. They are usually
performed on patients who suffer from severe osteoarthritis, or have suffered irreparable
hip fractures or aseptic necrosis (a condition in which the bone of the femoral head has
failed). THRs are most commonly received by elderly patients with the majority aged
between 50 and 60. The operation is also now being performed on younger patients [51]
and even animals [69].
In a human the hip, shown in Figure 2.1, consists of the hip socket or acetabulum,
a cup-shaped bone in the pelvis, and the spherical head of the femur, the femoral head.
Similarly a THR consists of two components, the femoral and acetabular components
in Figure 2.2, each corresponding to a part of the damaged hip interface. The femoral
Figure 2.1: Region of interest of a healthy hip. 1) Acetabulum, the large cup-shaped
structure on the lateral surface of the pelvis. 2) Femoral head which articulates within
the acetabulum
Figure 2.2: A total hip replacement. 1) Acetabular component, also called acetabular cup,
usually made from polyethylene. Of particular note is the elliptical metallic wire rim. 2)
Metal femoral head which articulates within the acetabular cup
component is typically made from titanium or cobalt-chromium alloy, while the acetabular
component is often made from ultra-high molecular weight polyethylene with the option
of a metal backing made from titanium or cobalt-chromium.
The THR operation begins with the patient anaesthetised and the surgeon makes an
incision over the hip joint. The soft tissue, the ligaments and muscles, are divided to allow
Figure 2.3: Region of interest of a total hip replacement. 1) Acetabular rim wire that
marks the opening of the acetabular cup. 2) Edge marker used by Charnley to measure
acetabular wear. 3) Femoral head, the spherical component of the total hip replacement.
4) Neck of the prosthesis. 5)Stem of the prosthesis
access to the bones of the hip joint. The femoral head is dislocated from the acetabulum
and sawn off. A canal is created within the femur by a rasp for the stem of the prosthesis to
enter. The spherical ball that acts as the femoral head is attached to the stem component.
The neck of the prosthesis is typically tapered to receive the femoral head and thus the
surgeon has several options as to where to place the femoral head on the neck. Cartilage
is removed from the acetabulum and the acetabulum is reamed into a hemispherical shape
to fit the shell of the acetabular component. Trial components are used to ensure the
components are the correct size for the patient. Once this is confirmed the acetabular
component is inserted into place. Once both components have been inserted the surgeon
replaces the bone and the soft tissues. The incision is sealed and the operation is concluded.
There are three categories of fixation: cementless, cemented and hybrid. Cementless
THR is preferred for younger, more active patients as the activities they typically participate in could put sufficient stress on the hip joint to cause cement to fracture. Cemented
THR is preferred for less active patients but remains the most popular method of fixation.
These are not hard and fast rules however, each surgeon has his or her own opinion as to
which method of fixation gives the best results.
In a cemented THR the acetabular cup and the femoral stem are cushioned to the
9
Figure 2.4: Diagram of a cemented THR

skeleton with a layer of polymethylmethacrylate bone cement. The bone cement fills
completely the space between the skeleton and the surface of the prosthesis, as shown
in Figure 2.4. The acetabular cup may be metal backed to allow the changing of worn
polyethylene components without disrupting the cement-bone interface.
In cementless THR the acetabular cup is press-fitted, held in place by the tightness
of fit and screws. The acetabular cup of a cementless prosthesis must always be metal
backed to aid in the press-fitting, as shown in Figure 2.5(a). Bone tissue does not react
well to direct contact with polyethylene or ceramic, and thus without a metal backing
would loosen quickly. The surface of the metal backing and femoral stem are coated to
aid fixation. Some have a porous coating, either a wire mesh or small metallic balls to
act as a grain (shown in Figure 2.5(b) and 2.5(c)), to allow the bone to regrow around
10
the prosthesis and thus achieve a stable fixation. Hydroxyapatite [38] is also a popular
coating for cementless THR stems as it acquires very strong bonds with living bone. Some
surgeons additionally use a hybrid technique [44] where one component is cemented while
the other is cementless. The exact fixation method used depends entirely on the surgeons
preferences and the patients requirements.
Conventional femoral heads range from 22mm to 32mm in diameter, though larger sizes
up to 54mm are not unheard of. Acetabular cup outer diameters (including any cement
spacers) range from outer diameters of 46mm to 64mm. The Zimmer Collarless Polished
Taper (CPT) system [23] supports femoral head diameters of 22mm, 26mm, 28mm, 32mm,
36mm and 40mm. The femoral component for primary total hip replacement comes in 5
sizes independent of the femoral head size, with differing stem lengths, neck length and
head/neck offset and it also has models with extended offsets and smaller stems for smaller
patients. Zimmers ZCA cups [24] range in diameter from 48mm to 61mm.
The CPT system and ZCA cup are but one of many different implant combinations.
Furnes et al. [36] reported that there were over 200 different implant combinations used in
Norway in 1996. Within the UK Murray et al. [67] noted that 19 different companies were
manufacturing 62 different models of implant manufactured for primary THRs. Half of
those were introduced within the five years of the publication date and yet merely 30% had
results published in peer-reviewed journals. As numerous new models are patented every
(a)
(b)
(c)
Figure 2.5: Components of a cementless THR include (a) the acetabular cup consisting of
an outer shell with porous coating and screw to allow fixation in bone and a polyethylene
liner (b) porous coating of the stem of the femoral component (c) bonding of the porous
coating to the bone
11
year it is essential to obtain evaluation of their survival and the economic consequences of
their use.
2.2
Failure of THRs
Unfortunately THRs do not last forever. Failure of a total hip replacement is defined
as when the prosthesis becomes ineffective, typically when the patient experiences the
symptoms of increasing pain and loss of motion. The typical lifespan of a prosthesis is
10-20 years [100] but there are still some cases that fail earlier. Similarly there are some
cases where the hip replacement has survived for more than its expected lifespan.
When a total hip replacement fails, revision surgery is needed. Revision hip surgery
involves the removal of the previous prosthesis, the cement (if applicable), the surrounding
tissue and damaged bone before a new prosthesis can be inserted. Revision prostheses
typically have to replace a lot more tissue. With over 5,000 revision THRs per annum
within the United Kingdom [51] this is an economic drain and distressing for the patient.
Additionally revisions have a higher rate of infection and poorer performance [8]. A young
patient who receives a total hip replacement could expect a revision as much as every 10-15
years of their life.
While some THRs are revised due to infection of the surrounding tissue it is aseptic
loosening (loosening without infection) of the prosthesis that is the chief cause of THR
failure. The Swedish Hip Register [64] notes that 75.3% of revisions in Sweden between
1979-2000 were due to aseptic loosening. Ilchmann et al. [52] stated that there was a definite relationship between loosening, migration and linear acetabular wear. Radiolucency,
a high rate of migration or a high rate of wear indicate that a prosthesis is likely to fail.
Loosening can be caused by osteolysis, a condition that affects bone volume. Bone is
constantly lost, or resorbed, and replenished continuously. The ratio of bone resorption to
bone formation controls the volume of bone present at any particular time. In the young
the level of formation should exceed that of resorption, while in the elderly the level of
resorption exceeds that of formation. Osteolysis is a condition that causes bone resorption
to exceed bone formation, causing a loss of bone volume. This leads to faster loosening
around the prosthesis-bone interface and the need for more revisions.
12
2.2.1
Radiolucency
There is no general definition of when an implant is considered loose. Loosening can

manifest itself on radiographs by radiolucent areas between the prosthesis and bone on
radiographs, in other words a cavity between the prosthesis and bone as seen in Figure 2.9.
The prosthesis may move within this cavity and movement causes pain and stiffness for the
patient. In order to assess loosening Gruen [41] created seven delineated sections around
the femoral component, shown in Figure 2.6(a). These were used for zonal evaluation of
looseness. Radiolucency of between 1-2mm on the film is considered normal, while 2mm
of radiolucency in one of these zones is evidence of loosening. Gruen identified four modes
of failure for THRs and noticed radiolucent lines common across each mode of failure.
These are summarised in Table 2.1 and diagrams of each mode are shown in Figure 2.7.
DeLee and Charnley [27] created 3 zones for zonal assessment of loosening of the
acetabular component, as shown in 2.6(b). The cup is considered loose if there is 2mm
or more radiolucency in all three DeLee zones, or if progressive radiolucency develops in
zones 1 and 2. 2mm or more radiolucency limited to one or two zones merely indicates a
possibility of loosening.
It should be noted that some radiolucency is normal in THR radiographs. Age related
(a)
(b)
Figure 2.6: (a) Seven Gruen zones around the femoral component (b) Three DeLee zones
around the acetabular component. Image taken from Gruen et al.[41].
13
Mode
Mechanism
Cause
Findings
IA
Pistoning Behaviour
Subsidence of stem within
Radiolucent Line between stem and
cement
cement in Gruen Zones 1 and 2 Distal Cement Fracture

Stem displaced distally in cement
IB
Pistoning Behaviour
Subsidence of cement man-
Radiolucent Line in all 7 zones
tle and stem within bone

II
Medial stem pivot
Lack of supermedial and

infrolateral
cement
Medial migration, proximal stem
sup-
port
Lateral migration distal tip
Cement fracture zones 2 and 6
III
Celvar Pivot
Medial and Lateral toggle
Sclerosis and thickening of the bone
of distal stem
at stem tip
Hang up of stem collar on
Radiolucent lines in zones 4 and 5
medial cortex
Windscreen wiper reaction
at distal stem
IV
Cantilever Bending
Loss of proximal cement
Stem crack or fracture
support leaving distal stem

fixed
Radiolucent lines in zones 1 and 2,
6 and 7
Table 2.1: Modes of Cemented Femoral Stem Loosening
Figure 2.7: Visualisation of the modes of failure. Image taken from Gruen et al.[41].
expansion of the femoral canal and cortical thinning give the appearance of progressively
widening radiolucency. Clinical decisions made with regards to radiolucency can vary from
surgeon to surgeon.
14
Figure 2.8: A fractured stem
Figure 2.9: Radiolucency of the femoral stem
2.2.2
Migration
A large change in the position of the femoral or acetabular component is another indication
of loosening. Termed migration this is the change of the bone-implant distance projected
onto the radiographic film. However some small amount of migration is to be expected
in THR. A prosthesis is considered to be loose when total migration exceeds 5mm [12].
Early migration is thought to give indication of an increased risk of failure. Krismer et
al. [60] found that a change in the prosthesis to bone interface of 1mm or higher within
the first 2 years after implantation indicates a high risk of loosening later on in the life of
the prosthesis.
15
2.2.3
Acetabular Wear
McKellop suggests that osteolysis in THR patients is due to the creation of submicron
particles of polyethylene [66]. Wear of the THR through adhesion, abrasion and fatigue
produces submicron particles, known as wear particles. Potential particles generated include polyethylene, metal (or ceramic depending on the material used) and cement particles. These particles are very small, 90% are less than 10 m, but the quantity of these
particles is of the order of billions or trillions. Some of these particulate debris can penetrate the interface between prosthesis and bone, stimulating the cellular response leading
to osteolysis [47] and loosening of the hip. Recent studies maintain this is a major cause of
failure. For example out of 100 prostheses 2 stems were revised due to osteolysis and loosening, while 16 cups were revised due to osteolysis induced loosening [51]. It was originally
asserted that the bone cement particles caused the lytic process and this brought about
the invention of cementless total hip replacement. Sadly femoral lysis is more prominent
among the cementless components, suggesting cement may actually protect against lysis
[103]. Polyethylene from the acetabular cup is the major source of these particles, being the
softest of the materials used in THR. Thus predicting the long-term trends of acetabular
wear allows identification of patients who could be at risk of accelerated osteolysis.
The literature refers to two kinds of wear: linear and volumetric wear. Linear wear
is the maximum depth of the largest pit on the acetabular cup surface. It is typically
quoted in mm or as a rate of wear over time (such as mm/year). Volumetric wear is
an estimation of the volume of lost polyethylene. This varies with the penetration rate
and the area of contact. By assuming that the wear-tract is cylindrical it is possible to
estimate volumetric wear given linear wear by the following equation [62]
v = r2 w
(2.1)
where v is the volume of debris resultant from the wear (measured in mm3 ), r is the known
radius of the femoral head (in mm) and w is the linear wear that has been measured (also
in mm). Equation 2.1 does not take into account direction of wear. Kabo et al. [54] found
that the actual wear obtained from retrieval studies was 53% of that estimated using
Equation 2.1. The following equation [58] takes into account the direction of the linear
16
wear
v = r2 w
1 + sin()
2
(2.2)
where is the angle between the major axis of the acetabular rim and wear line. The
wear-tract is assumed to be cylindrical. It is assumed that the cylinder represents a
wear tract, and thus that all of the calculated volume represents the loss of polyethylene
particles. Penetration can also be caused by deformations of the polyethelyne. This
irreversible polyethylene deformation, known as creep, does not cause loss of polyethylene.
Polyethylene flows away from the area of high pressure and the femoral head penetrates
the acetabular cup in the direction of the applied force. This can in theory lead to
overestimating acetabular wear [19]. However polyethylene deformation has been found to
be negligible after 12-18 months. Initially deformation was a significant clinical problem,
leading the cup failures due to fracturing or complete penetration but advances in the
manufacture of polyethylene have all but eliminated it as a cause of failure.
Only large amounts of polyethylene wear are readily apparent to the casual human
observer, such as when the femoral head breaks through the acetabular component completely. This is unsurprising given that the annual linear wear rate has been found experimentally to range between 0.09-0.25mm/year [58]. A system for measuring acetabular
wear should be accurate enough to measure wear rates of this magnitude. The greater
the accuracy of the system the sooner evaluations of new prosthesis combinations can be
obtained, thus reducing the number of implantations using inferior designs.
Attempts to reduce the rate of polyethylene wear have led to the creation of more
durable acetabular cups. Cates et al. [14] noted that metal backing in cemented THRs
was intended to stiffen the polyethylene and produce a more even distribution of stress
but in fact led to an increased rate of wear. More success has been met by crosslinking polyethylene carbon-hydrogen chains. This has been shown to decrease the rate of
wear [66]. Ceramic components manufactured from alumina and zirconia are smoother,
harder and in the laboratory have shown a reduced rate of volumetric wear [39]. However
there have been problems with ceramic THRs - such as the recall of St. Gobain Desmarquests ceramic femoral heads due to their high fracture rate after a manufacturing process
17
change in 1998.
2.3
Radiography
Regardless of fixation or material all THRs are assessed by radiographic analysis. Computerised Tomography is hampered by surface scatter from metal implants, preventing it from
being used to assess THR. Looney et al. [63] have created specialist software to correct for
this scatter, but this is not used in standard clinical practice and has only been applied to
measuring osteolysis and correlating with acetabular wear measured in radiographs. Other
imaging techniques have not been applied to assessing total hip replacements. Magnetic
Resonance Imaging is similarly peturbed by image artifacts. Ultrasound does not provide
sufficient image quality to allow assessment of a THR.
The acquisition of a radiograph presents certain assumptions and limitations and thus
it is necessary to review the process of obtaining an anteroposterior (AP) radiograph of
the pelvic region [4, 97]. Figure 2.10 shows a simplified diagram of a system for obtaining
THR radiographs.
The x-ray system in Figure 2.10 is an example of perspective projection. It is similar to
that of a camera except that the image plane or film precedes the object in a perspective
projection model.
The radiologist aims the centre line of the radiation (called the aim line and indicated
via a beam of visible light) at the centre of the intended region of interest. In a radiograph
of the hip the projection centre is usually near the contour of the symphysis pubis (see
Figure 2.11), a joint of two large bones at the front of the pelvis. Photons are emitted
from the x-ray tube into the region of interest, such as the pelvic region. The filter limits
harmful low kilovoltage x-rays that would not penetrate a patient. They would only serve
to increase the dose absorbed by the patient. The rectangular collimators are placed so
as to ensure that only the desired volume of the patient is irradiated.
There are minor differences in intensity on the film from the peripheral rays of the
beam due to the different distances they have to travel. It is assumed that the initial
intensity is constant throughout the film. The x-ray tube emits photons that enter the
patient and can be scattered, absorbed or transmitted without interaction. Scattered
18
photons are removed by means of an anti-scatter device, such as an air gap or a lead grid,
that is placed between patient and film.
As the beam moves through the tissue of the patient it should emerge with reduced
intensity. The amount that the intensity is reduced by depends on the type and thickness
of tissue that it passed through. This is known as differential attenuation and is due to
the varying thickness of anatomical structures, the molecular number and physical density.
Figure 2.10: Obtaining a THR radiograph through a simplified x-ray imaging system
Figure 2.11: The symphysis pubis
19
The variations of intensities should reflect the layout and anatomy of the tissues that the
beam traveled through. These variations are known as subject contrast.
It is possible for the radiologist to control contrast, either by adding a contrast agent,
or by altering the x-ray tube kilovoltage. Increasing the kilovoltage reduces the contrast,
whereas decreasing it increases the contrast. The attenuation properties of bone compared
to surrounding soft tissues allows for good contrast over a range of kilovoltages.
Some x-rays exit the patient and come into contact with the film. Silver halides,
chemicals with atoms from the halogen elements such as silver iodo-bromide which is the
most commonly used halide in the medical field, are embedded in the film emulsion and
as a result of contact with the photons chemical reduction of the silver halide occurs. A
photon is absorbed in the crystal and causes the halide ion to release an electron. This
electron moves inside the crystal but inevitably becomes lodged in a low energy electron
trap near the surface of the crystal. Areas where this trapping occurs collect more and more
electrons, maintaining a negative charge, attracting the positive silver ions. Silver ions so
drawn to the negative charge gain an electron that neutralizes their charge and forms silver
ions on the film. However at this stage the image is invisible to the human eye, known as
the latent image. Processing the film enables the entire crystal of development centers to
be completely reduced to silver via the introduction of a developing fluid. Resultant x-ray
films are then often digitized, typically on a flatbed scanner with backlighting.
The distance of the x-ray source from the film is known as the Film Focus Distance
(FFD). This is not always recorded by the radiologist but has been found experimentally
to be between 900 mm and 1300 mm [12, 59], though it is typically quoted as being
1000mm [68]. The distance from the centre of the femoral head, or the prosthesis reference
point, to the film has similarly been determined to be between 180 mm and 270 mm. This
is referred to as the Object to Film Distance (OFD). It should also be noted that two
forms of rotation occur during the process of obtaining a radiograph. Interior rotation is
the rotation of the film plane with respect to the beam centre, while exterior rotation is
the rotation of the pelvis relative to the film plane.
A key point in the process is that a radiograph is a two-dimensional projection of the
three dimensional distribution of the x-ray attenuating properties of the tissue. Another
point that is important to make is that correction for magnification must occur when
20
measuring from radiographs. 1mm on the radiographic film does not correspond to 1mm
of the structure of interest. In the case of a THR the diameter of the femoral head can be
used to correct for this magnification error. Lastly objects such as the spherical portion
of the femoral head manifest as an ellipse within the plane of radiographs [59].
It is often desirable to obtain in addition to the AP radiograph a mediolateral (ML)
radiograph. Ideally the two views should taken simultaneously at exposure, but such
stereo-imaging equipment can prove costly. ML radiographs can be of poor quality as
they are more difficult to standardise [68] and positioning the patient is often difficult due
to discomfort. Additionally contralateral structures can be superimposed on the hip on
an ML view. Thus ML data is not always available to the clinician.
2.4
Summary
In this chapter the rationale behind the THR operation was explained and an overview of
the THR was provided. The major cause of THR failure was identified as aseptic loosening
due to osteolysis caused by wear of the polyethylene acetabular component. Therefore a
method to assess acetabular wear is desirable. This system must be capable of analyzing
THR radiographs with sub-millimetre accuracy. However when taking a radiograph there
are several important variables that have already been set by the radiographer, including:
the FFD, which is not always recorded at exposure
the OFD, similarly not always recorded at exposure
the exterior rotation - relative orientation and position of the pelvis at exposure
interior rotation - relative orientation and position of the film at exposure
the diameter of the femoral head
the position of the centre beam and its intersection with the film
existence of ML radiograph
In the next chapter existing methodologies for measuring wear are examined. In addition to the variables noted above each method of assessment will have its own inherent
variables depending on what assumptions are made and what landmarks are used.
21
Chapter 3
Related Work
This chapter briefly reviews the literature on methods for radiographic assessment of total
hip replacements. There is a surprisingly large amount of literature detailing means of
radiographic assessment ranging from labour-intensive methods requiring manual analysis
of the radiograph film to fully or semi-automated computer applications, allowing the user
to make measurements with the aid of image processing techniques. This related work
is divided into four broad categories: methods that were intended to allow assessment of
2-dimensional (2D) acetabular wear, assessment of 3-dimensional (3D) acetabular wear,
assessment of 2D migration and 3D migration. The migration methods included in this
review can be used to measure wear but this was not the primary motivation behind their
creation.
It is also necessary to classify each of the methods as uni-radiographic or duo-radiographic.
A uni-radiographic method requires analysis on solely one radiograph - the most recently
obtained radiograph. A duo-radiographic method requires two radiographs, typically the
postoperative radiograph and the latest follow-up radiograph.
It is difficult to draw direct comparisons between the stated accuracies or repeatabilities of the methods mentioned here as there is often little indication as to how these
results were obtained. Some methods state a mean and standard deviation obtained by
examination of paired measurements while others have an error obtained by comparing
the observed measurement against a gold standard such as a hand-annotated image or the
shadowgraphic technique - a method for assessing the volume of polyethylene lost from a
22
retrieved prosthesis. Some of the units correspond to linear wear while others correspond
to volumetric wear with no clear indication of how to convert between them.
3.1
2D Wear
Methods for assessing 2D wear typically concern themselves with measuring linear wear
solely within the AP view, as this is where the majority of wear manifests itself [50]. The
observed value of linear wear is then used to estimate volumetric wear by Equation 2.1.
Existing work in this area is directly relevant to the method described in this thesis since
it too assesses 2D acetabular wear.
Initial studies into measuring acetabular wear were conducted by Charnley et al. [15]
using a semicircular radiographic marker introduced ten years prior to their initial study.
This wire was placed on the periphery of the acetabular cup. The shortest distance from
the femoral head to the acetabular cup edge wire marker (thinness) was subtracted from
the largest distance (thickness) in the non-weight bearing zone, as shown in Figure 3.1.
The resultant value was then halved. This uniradiographic method was only valid if
the maximum wear occurred in the weight-bearing area. Its main attraction was that it
required only one radiograph and thus measurements could be obtained immediately from
the latest radiograph. Given that radiographs use silver halides it is often economically
desirable to recycle the silver rather than keep postoperative radiographs indefinitely.
Charnley et al. also proposed a duo-radiographic method [16] that compared the thickness
of the postoperative radiograph with that of the current follow-up radiograph.
Charnleys initial studies estimated the mean 2D wear to be from 0.1 mm/year to 0.2
mm/year after 5 to 6 years of operation. At 10 years this was found to be 0.13 mm/year
using the uniradiographic method. Later work using the duo-radiographic method Charnley stated the average to be 0.15 mm/year at 10 years, with 68% of cases being equal to
or less than 0.15 mm/year. The average of these 68% cases was 0.13mm/year while the
remaining 32% of cases averaged at 0.32mm/year. The most common direction of wear
was described as having an upwards-and-outwards direction.
Charnleys methods for measuring acetabular wear were criticised by Clarke et al. [18],
who claimed there were too many variables inherent in the acquisition of THR radiographs
23
to accurately measure wear. Such variables included rotation of the wire marker out of the
coronal plane, or the distance of the reference points from the beam centre. They concluded
that the duo-radiographic technique was more accurate than the uniradiographic method,
but was still unreliable. Clarke et al. found the magnitude of error of using both methods
to be larger than the expected amount of wear. However Griffith et al. [40] stated that
the duo-radiographic technique could give consistent wear measurements provided the aim
line was centered on the symphysis pubis and that the wear marker was not rotated more
than 10 from the coronal plane. Measurements could differ by up to 0.2mm depending
on the wire markers alignment with the coronal plane. However Griffith et al. also found
that the mean wear observed in the new measurements obtained from a dataset with an
average age of 8.3 years was half of what Charnley had originally reported at the 10 year
stage, at 0.07mm/year.
Wroblewski [99] validated Charnleys duo-radiographic method by assessing prosthetic
wear in retrieval studies and indicated that there was strong correlation between Charnleys
measurements and the actual wear noted in 22 retrieval specimens. Using Charnleys duoradiographic technique a mean wear of 0.21 mm/year was measured while the actual mean
wear was 0.19 mm/year.
A duo-radiographic method was proposed by Livermore et al. [62]. This method for
measuring acetabular wear used a transparent overlay with a set of concentric circles with
Figure 3.1: The narrowest and widest areas in the weight-bearing area of the prosthesis,
also known as the thinness and thickness
24
1mm increments of radial length. The transparency was superimposed on the radiograph
film to determine the centre of the femoral head. Magnification was corrected for by the
known diameter of the femoral head. Then by means of a compass and caliper the shortest
distance to the acetabular cup cement interface from the centre of the femoral head was
determined on the latest follow up radiograph. This was identified as the largest wear tract.
The angle of wear was determined as the angle of the centre of the femoral head to the area
of maximum wear relative to the tangent of a line drawn through the caudal points of the
ischial tuberosities (see Figure 3.2). The thickness at the same position on the acetabular
cup was measured on the postoperative radiograph and the difference between the two
measurements was taken as a measure of wear. Livermore also first noted the relationship
between volumetric wear and the femoral head radius and recommended using the smallest
femoral head possible is desirable. Paired measurements of using Livermores method were
found by Martell et al. [65] to have a standard deviation of 0.72mm when conducted by
the same observer.
Livermore et al.s method remains the most widely used method for assessing acetabular wear[72], no doubt due to its low setup cost. It requires no special equipment and
can be performed directly on the radiographic film immediately. The repeatability of the
method is doubtless low as it is a very labour intensive method of measurement. Cowell [25] noted there is considerable intrapersonal and interpersonal variability when two
surgeons draw the same line on a radiograph, much less perform a method of this level of
labour intensity. Callaghan et al. [13] extended Livermore et al.s technique by additionally
finding the direction of wear by superimposing the most comparable first postoperative
film on top of the final followup film and marking the direction of wear by pencil.
Jones et al. [53, 43] created the first computer-assisted method for assessing total
hip replacements. The Manchester X-Ray Image Analysis system (MAXIMA) was an
interactive tool that aided in the location of reference points on the prosthesis, including
the centre of the femoral head, midline of the stem and the tip of stem. Beyond results
from an experimental model no further data was ever published about the MAXIMA
technique and it was not widely adopted. It should be noted that the method was labour
intensive, dependent on a series of interactively chosen landmark points which initialised
a local search using a custom edge enhancement filter to locate the contour of the femoral
25
head and stem.

Shaver et al. [87] created a duo-radiographic image analysis program for the determination of linear wear. The user selected three points, on the medial, lateral and central
points of the elliptical portion of the femoral head, as shown in Figure 3.3. An initial estimate of the femoral head centre was obtained, equi-distant to the three selected points.
Sampling rays emanated from this centre point estimate at intervals of 0.5 . These rays
were constrained to within the elliptical portion of the femoral head by two landmark
points. Points where the surrounding grey-scale profile did not match the expected pattern of grey-scale intensity were discounted. This resulted in a locus of points clustered
around the contour of the femoral head. An ellipse was fitted through these points using
the least-squares ellipse fitting method [34, 73]. This procedure was then repeated to
find the contour of the metal backed acetabular cup. Shaver used both the change of the
displacement of the femoral head centre relative to the acetabular cup centre over time
and the separation changes between corresponding points on the elliptical arcs. Shaver
stated that the latter method was more precise. On radiographic data where the true wear
was known Shavers method showed a mean error of 3.6% compared to Livermore et al.s
method which showed a mean error of 23.1%.
The two landmark points on the start and end of the elliptical portion of the femoral
head prevent erroneous points on the non-elliptical portion skewing the least squares ellipse
Figure 3.2: The ischial tuberosities and a line through the caudal points are used to
determine the angle of wear.
26
fit. The use of the custom edge filter is an example of a bottom-up strategy. It is interesting
to note the method could easily be applied to ML radiographs but was not.
Chen et al. [17] created a duo-radiographic method whereby given the known radii
of the femoral head and acetabular cup, the observed radii on the x-ray and the known
FFD of each radiograph the depth of wear can be estimated without the need for an ML
radiograph. Circles were fitted to the femoral head and acetabular cup contours and their
observed radii were compared to the actual radii. The distance of the x-ray source to the
centre of the cup to the film is estimated and then applied to the centre of the femoral
head. This distance was used to estimate the third radii for the femoral head. If this was
smaller than the observed femoral head radius the femoral head was thought to have worn
toward the x-ray source whereas if it was larger than the observed radius it was assumed
to have worn away from the x-ray source. Further calculations allowed the wear depth to
be obtained for the z-axis. This method assumed that the FFD is recorded at exposure
and thus no attempt is made to measure the propagation of errors in recording the FFD,
much less errors when the FFD must be approximated using the maximum, minimum
and mean values of Krismer et al. [59]. Chen et al. fitted circles instead of ellipses to the
contours of the acetabular cup and femoral head.
Hatfield et al. [45] proposed two semi-automated uni-radiographic computerised methods for measuring acetabular wear. By user-selection of the medial, lateral and central
Figure 3.3: Medial (m), lateral (l) and central points (c) with the initial estimate of the
centre of the femoral head determined by a circle fit
27
points of the femoral head (see Figure 3.3) an initial circle was fitted. Sampling rays emanated from the centre of this circle, constrained to remain within the lateral and medial
points at 1 intervals. By means of the same edge enhancement filter that Jones et al.
had employed the point of maximum gradient on each profile was selected. If the mean
intensity on both sides of the profile did not meet prior expectations of the edge of the
femoral head contour it was excluded and the next highest gradient point was considered.
Once all the points had been located the user manually selected outlying points to be
removed. A second circle was fitted through all the remaining points and any point with
a distance of more than three standard deviations from the circle contour was removed.
A third and final fit would then localize the femoral head contour.
The acetabular cup edge wire was detected in a similar manner. In addition to the
lateral, central and medial points an additional two points on the edge wire contour were
required to fit an ellipse. A different edge filter was used to locate points on the acetabular
edge wire contour. However the first method used a circle to localize the acetabular cup
edge wire, while the second method used an ellipse. The two methods were compared
with the shadowgraphic technique and it was found that the acetabular cup centre was
distorted if the wire marker was not circular, and thus not aligned with the coronal plane.
Hatfield et al. stated that not even using an ellipse to localize the wire marker contour
would account for this. Incorrect measurements could be made if the wear did not occur
in the plane of the AP radiographs. It is interesting to note that this method had to
account for a radiopaque structure known as the Mexican Hat, a wire mesh used to plug
the gaps in the pelvis during surgery. This was handled by allowing the user to exclude
points found on the Mexican Hat. In order to fully automate measuring acetabular wear
it would be necessary to detect and remove erroneous points without user intervention.
Eggli et al. [31] created a semi-automated uni-radiographic method for measuring wear
which again required the interactive selection of the medial, lateral and central points of
the femoral head to allow a circle to be fitted to the femoral head contour. The centre
of the cup was specified by points corresponding to the ends of the major axis of the
elliptical acetabular rim. The Euclidean distance between the two centroids was taken as
a measure of wear, with magnification corrected by the known diameter of the femoral
head. Volumetric wear was calculated using the Euclidean distance as per Equation 2.1.
28
The annotation of the major axis of the acetabular rim was most likely very subjective,
and to compensate for this Eggli used radiographs scanned at 2000 dpi and magnified 10
times. Such a high resolution is not typically used in clinical practice. It should also be
noted that Eggli made no attempt to correct for the eccentricity error of the acetabular
rim ellipse. Eggli further confounded matters by requiring location of the teardrop. This
is a teardrop-shaped pattern that exists on both sides of the pelvis and is visible in AP
radiographs. This teardrop represents the summation of various osseous planes of the
acetabulum. In this instance it is used to define a co-ordinate system for measuring wear.
Redhead et al. [74, 75]used active shape models (ASMs) [22], a method combining a
flexible point distribution model of shape with several grey-level models learned from a
set of training images. In this instance the ASMs were used to identify the position of the
femur, pelvis, stem and acetabular rim in radiographs. The femur and stem were located
and point-to-boundary error was measured. This gave mean error of 0.7mm whilst the
femur, stem and pelvis were located with an error of 2.4mm. This is unsurprising given
the variation of the pelvis and its weak edges on radiographs. The ASM was not directly
applied to any form of THR assessment, though its application to measuring acetabular
wear was briefly considered, perhaps due to the high error. Redheads work does highlight
the difficulty in segmenting areas of the pelvis in radiographs.
3.2
3D Wear
Methods for assessing 3D wear consider both the AP and ML view and obtain a measure
of linear wear for each view. The two views are obtained at different intervals in time and
often require repositioning the patient, thus no calibration occurs. Several of these methods
can also measure 2D wear in the absence of ML radiographs but what differentiates them
from the wear methods is that the primary motivation of the methods here is to use both
views.
Martell et al. [65] created a duo-radiographic method for computer-assisted vector
wear analysis of metal backed prostheses using digitized AP and ML radiographs. This
method could additionally measure wear solely in the plane of the AP if necessary. Two
points were placed on the caudal most parts of the ischial tuberosities, an area of the
29
lower part of the pelvis shown in Figure 3.2. Using custom-designed variations on the
Sobel edge filter circles were fitted to the femoral head and acetabular cup via Hough
transform. The user was given the option to override erroneous fits by limiting the region
of interest or by choosing three points on the contour of the femoral head or acetabular
cup. In an experiment examining paired measurements using the method the repeatability
was given as 0.004mm for paired measurements by the same observer or 0.06mm for
paired measurements by differing observers.
A vector is created for each image. The origin of this vector is the centre of the acetabular cup. Wear is the distance between the vector of the postoperative radiograph
from the vector of the current radiograph. Knudsen et al. [57] compared Martells vector wear method and found that it had a higher standard deviation of paired measurements (0.332mm) than computerised measurements obtained by measuring the thickness
(0.238mm or 0.267mm).
Vector wear was defined as the centre of the femoral head on the postoperative radiograph through to the centre of the femoral head on the most recent radiograph. The
centre of the acetabular cup on one radiograph is superimposed on the other to provide a
reference point when determining the amount of wear.
Devane et al. [28] attempted to measure 3D wear through the use of AP and ML radiographs, taken independently, and to estimate the volume of polyethylene debris created.
The beam centre and the FFD were assumed to be known by the user. Three points were
placed on the opening of the acetabular cup in the AP radiograph to which an model of
the acetabular rim, consisting of two concentric ellipses corresponding to the inner and
outer edges of the wire, was fitted. Additionally circles were fitted to 5 to 12 points placed
on the contour of the femoral head and the acetabular cup in both AP and ML radiographs. Circles were fitted to these points and 2D wear for each radiograph was calculated
as the change in the position between the two circles over time on followup radiographs.
3D models of the cup and head were created for corresponding AP and ML radiographs
using the manufacturers specifications. Figure 3.5 shows two example models. The tilt
and anteversion of the cup is then calculated with respect to the cup and used to rotate
the models to a standard frontal view to account for exterior rotation. Frontal views created from post operative and followup radiographs could be compared allowing wear to
30
be measured as the difference between the displacements of the femoral head relative to
the acetabular component. Volumetric calculations are made as per Equation 2.1. The
accuracy of the technique was obtained by volumetric measurements on 10 sets of AP and
ML radiographs of known volumetric wear. The accuracy of these calculations was 0.411
cm3 .
Polyware [29], the commercial version of Devane et al.s method, used a custom edge
detection filter to correct annotations, reducing the number of landmarks points required
from 5 to 3. Figure 3.4 shows the edge detection correcting three poorly chosen landmark
points. Despite this Polyware remains a labour-intensive method, even with the aid of edge
filters to automate femoral head and acetabular cup contours. Clinicians have reported
finding Polywares computer interface complex [72]. Estimating the beam centre on the
radiograph can be very problematic and if incorrect the resulting models are erroneous.
The edge detectors assume that the cup is metal backed and that there is sufficient contrast
between the femoral head and acetabular cup to allow both contours to be localized.
Polyware is used in clinics in the United States, but its cost can be prohibitive at $20,000
for the software [72].
Figure 3.4: Circle fitting in Polyware
31
Figure 3.5: Model created using Polyware showing the AP (left) and ML (right) views
3.3
2D Migration
Methods of migration measure the displacement of the prosthesis from the bone over time,
typically by selected reference points on the bone and on the prosthesis. When both the
migration of the femoral head and acetabular cup is known wear can be computed, thus
justifying the inclusion of these methods in this review. However many of the papers cited
herein imply that migration of the femoral head and acetabular cup are interchangeable.
What makes these works relevant to this thesis is that acetabular wear can be obtained
by measuring the femoral migration and acetabular migration and taking the difference
of the two.
Sutherland et al. [89] developed a method for measuring migration using the centre of
the acetabular rim and three reference points, as shown in Figure 3.6. The first of these is
the caudal most point of the adjacent teardrop. The remaining pair of points come from
the Kohler line, a line drawn along the medial aspect of the ilium and ischium on the AP
radiograph. The x-distance is found by drawing a line orthogonal to the Kohler line to
the centre of the acetabular rim, which is in itself found by drawing the major axis on the
radiographic film. The y-distance is found by a horizontal line from the caudal teardrop
and drawing a vertical line from the centre of the acetabular rim that contacts with the
horizontal line as shown in Figure 3.6.
Nunn et al. [68] created a method for measuring acetabular component migration
32
directly on the radiograph. A line is drawn passing through the caudal points of the
teardrops (Line T in Figure 3.7). The centre of the femoral head is found by using the
transparent overlays of Livermore et al.. A line that is orthogonal to Line T is then drawn
(Line P in Figure 3.7). The x-distance and y-distance from the femoral head centre to
the adjacent teardrop point is measured and corrected for magnification by means of the
known femoral head diameter. The accuracy of the technique was stated as being 3 mm,
thus the technique is only suited to measuring large amounts of migration in the absence of
wear. Dickob et al. [30] created an interactive image analysis program that used a similar
technique to Nunn et al. but rather than the femoral head centre the software used the
centre of the acetabular rim as a reference point, or the centre of the cups equatorial
plane if it was metal backed.
Each of these methods used reference points on the pelvis to measure the 2D migration. Selection of these points is highly subjective, resulting in high interpersonal and
intrapersonal variation. Additionally as noted in Chapter 2 the pelvis position is not identical in a series of radiographs, so systematic errors result from using radiographs with
disparate positions. These methods also imply that the centres of the femoral head and
acetabular cup are interchangeable. Nunn et al. [68] even stated that they were aware this
erroneously assumes that there is no wear.
Figure 3.6: Migration measured by Sutherland et al.s method. m1 is the caudal most
point of the adjacent teardrop. m2 is the ischium on the Kohler line. m3 is the ilium on
the Kohler line.
33
Krismer et al. created the Einzil-Bild-Roentgen-Analyse (EBRA) method[59] for determining migration of the THR. The foramina obturatoria were used as bony landmarks,
by placing a series of lines (t1 to t6 in Figure 3.8) on pelvic structures to define the
position and rotation of the pelvis. A locus of points was required to obtain an ellipse
Figure 3.7: Migration measured by Nunns method. T) A line passing through the caudal
points of the teardrop to measure y-distance. P) Orthogonal to Line T to measure xdistance.
Figure 3.8: EBRA method with reference lines placed on pelvis and annotated points on
both the femoral head and acetabular rim.
34
via least-squares fitting[34, 73] to localize either the femoral head, the edge wire of the
acetabular cup or the acetabular rim as an implant reference point. These were all placed
relative to the beam centre, which is assumed to be the midpoint of a line connecting
the film centre and the cranial end of the symphisis pubis. Using the extremal values of
the FFD (900mm to 1300mm), the known radius of the femoral head and its apparent
radius on the image, an upper and lower estimate of the FFD of the image was obtained.
From both of these values a 3D implant reference point, the x and y reference planes and
the distances between the reference point and the planes were computed and averaged to
give a final 2D distance of migration. A comparability algorithm examined the distance
between the tangents on a series of radiographs and determined pairs of radiographs that
are comparable in terms of interior and exterior rotation by a comparability limit typically
set to 2 or 3mm. Wear was indirectly measured as the difference between head and cup
migration or by the displacement of the femoral head centre relative to the acetabular cup
centre.
Unlike the previous method EBRA employed reference lines rather than reference
points. It also supported different types of prosthesis, allowing for the use of the metal
backed cup, the acetabular rim or the acetabular edge wire to measure migration. EBRA
was shown to have an accuracy of 1mm and while it improves on the previous techniques
could not measure early migration. It should only be used to measure migration later in
the lifespan of the prosthesis and due to the comparability algorithm requires a series
of at least 4 radiographs, with more radiographs increasing the likelihood of finding a
comparable pair. It should also be noted that some of the techniques used in EBRA have
been extended to measure stem tip migration. This incarnation of the EBRA method is
known as EBRA-FCA (Femoral Component Analysis) [5] though this is not relevant to
the discussion of measuring acetabular wear.
3.4
3D Migration
Roentgen Stereophotogrammetic Analysis (RSA) [56, 86] was created by Selvik et al. as
a general method for obtaining 3D measurements of radiographs that can be applied to
measuring migration of the prosthesis. This involves the insertion of tantalum markers
35
(also referred to as beads or balls) by means of an injector with a spring loaded piston (see
Figure 3.10). The beads have a diameter of either 0.5mm, 0.8mm or 1.0mm and are placed
into the greater and lesser femoral trochanters, tuber ischii, the roof of the acetabulum
and within the acetabular cup as shown in Figure 3.9. These markers can themselves
migrate small distances within the patient just after insertion so it is typical to wait 10-14
days before obtaining radiographs.
Unlike conventional radiography described above two radiographs were obtained with
simultaneous exposure. Two x-ray sources were positioned at approximately 1.6m above
the film at a 20 angle to the vertical (see Figure 3.11). A combined reference and calibration plate with its own tantalum markers at known positions was placed beneath the
patient. The two-dimensional distances between the images of the tantalum markers on
the two films provided the input data for computerized conversion to a three-dimensional
coordinate system using an interactive package. Thereafter, migrations and rotations of
the femoral head, stem and the acetabular cup could be calculated. The difference between
the acetabular cup and femoral head migration was again an indicator of acetabular wear.
Figure 3.9: An example radiograph with tantalum markers shown

Manual RSA measurements on the film were very time consuming and required a great
deal of user interaction. RSA Clinical Measurement System (RSA-CMS) [96] automatically detected and classified the tantalum markers in digitized RSA images via a variant of
the Hough circle-finder. It should be noted this is a fully automated method that improved
upon the accuracy of the manual RSA measurements, though user interaction could be
36
(a)
(b)
Figure 3.10: Insertion of tantalum markers into (a) the femur (b) the acetabulum
(a)
(b)
Figure 3.11: (a) Typical setup for obtaining an RSA radiograph. (b) Hip replacement
positioned at the intersection of both x-ray sources
required to correct for erroneous detections.
RSA has a high accuracy of measurement (10-250 m) and it is considered an invasive method for measurement. It cannot be used retrospectively, the markers must have
been implanted with the prosthesis. In some countries implantation of the tantalum markers must be approved by regulatory bodies. Costs of the stereo-imaging equipment are
37
prohibitive for wide-spread use. In 2000 the set up cost for RSA was stated as being approximately $90,000[72] with a cost of $1 per analysis. Additionally the tantalum markers
can become loose and migrate into the soft-tissue, becoming extra-osseous[61].
Model-based roentgen stereophogrammetry has been proposed as an alternative to the
implantation of tantalum markers. This research has focused on knee implants as the
metal implant can obscure the tantalum markers but this is also true of THRs with metal
backed acetabular cups. Thus this research is directly applicable and relevant to THRs.
Hoff et al.[46] proposed a method for estimating the pose of femoral and tibial total knee
replacement components in fluoroscopy images. This used a library of synthetic images of
knee replacement components generated from a 3-D model at known orientations. These
were used to determine the pose of novel fluroscopy images by performing a series of template matching operations and using the pose of the highest correlating synthetic image.
Sarevock et al.[84, 85] extended this method to use the 3-D model directly, eliminating
the need for the library of images.
Valstar et al. [93] initially proposed a method whereby the pose of the implant in
RSA images was calculated by minimizing the non-overlapping area of a CAD model of a
prosthesis with the detected prosthesis contour in a digital radiograph in a manner similar
to Sarevock et al. The contour in the image was detected by the Canny edge-detector and
interactively selected from the resultant edge map by a human user. The initial estimate
of the model parameters was supplied interactively by the user. The accuracy of this
method when used on radiographs of known pose was found to be considerably less than
that of marker-based RSA due to dimensional tolerances of the implant.
Kaptein et al.
[55] extended this RSA method, using both a CAD model and sev-
eral laser scanned reverse engineered models of the prosthesis with varying number of
constituent elements. Unreliable parts of the model contour were removed, thus meaning
that instead of minimizing the non-overlapping area between the CAD model and detected
contour the Euclidean distance between corresponding nodes on the model and contour
were minimized. It was found that the reverse engineered models provided more accurate
measurements than the CAD model, and for the reverse engineered model with the most
elements the maximum standard deviation of the error was found to be 0.06mm for xand y-translation and 0.14mm for out of plane z-translations. These results were obtained
38
from a phantom model and showed promised, but data for model-based RSA in a clinical
setting is not yet available.
3.5
Summary
This chapter has examined methods for radiographic analysis of THRs, ranging from
manual methods performed on the film of the radiograph to fully automated 3D methods.
All of the methods for measuring 2D and 3D wear required a level of manual labour
which introduced error and resulted in both intra- and inter-personal variation in the
measurements obtained. RSA-CMS is the only fully automated method for assessment of
migration and is often too costly and invasive to be used in standard clinical practice.
From this literature it is apparent that a method is required that must be easy to
use, not requiring more than a standard PC setup if it is to be used in standard clinical
practice. Automation seems a desirable quality. Too many of the methods require different
reference lines or reference points on the pelvis. The selection of these landmarks is highly
subjective and given the difficulties encountered by Active Shape Models in locating pelvic
structures it would be advisable for an automated method to avoid using these structures.
Such a method must also be capable of finding the correct solution in the presence of such
radiopaque debris as screws to aid fixation, or the Mexican Hat described in the work of
Hatfield et al. With these observations in mind Chapter 4 begins the process of developing
a system to measure acetabular wear by considering how to model both the shape and
appearance of the relevant contours, while avoiding any bony landmarks. Chapter 5 then
details a robust method of locating these contours in novel images using the model.
39
Chapter 4
Modelling
The method detailed in this thesis can be considered an example of model-based vision.
Using prior knowledge of shape and appearance the method can make a reliable estimate
of the position of the desired structures in novel data. This chapter explains the process of
modelling the shape and appearance of the femoral head and acetabular rim wire marker
contours. By using the distinct normalized first order grey level derivative the system can
locate points on the desired contours. These can allow the area of interest to be localized
by fitting an ellipse through the points. Chapter 5 expands on the notion of fitting to these
points while this chapter details the process of modelling both shape and appearance.
4.1
Overview
Modelling is referred to as a top-down design stratagem - starting at the high level, considering the application and refining to the lowest level. Thus initially the problem of locating
the desired structures is considered and the shape and appearance of these structures are
modelled. This model is then applied to novel data. By contrast in bottom-up design
components are then linked together to form larger components until a complete system
is created. Earlier computer vision applications applied bottom-up strategies where the
emphasis was on designing a general and theoretical method (such as an image filter) independent of any real dataset and then applying it to (among other applications) localising
the desired radiographs in THR radiographs. Examples of these alluded to in Chapter 3
include the work of Shaver et al. [87], using the Sobel edge filter, or Hatfield et al. [45]
40
and Jones et al.s [53] use of custom edge filters.
4.2
Previous Work
Shape models are a popular and powerful tool to aid in interpreting images. They encompass a variety of forms. Active Contours [6, 91] or snakes minimize an energy function to
converge on structures in image feature maps. Constraints can be placed upon an active
contour but tend to encourage rather than enforce a particular class of shapes and thus it
is difficult to represent the priori that the contours are elliptical. Additionally snakes are
sensitive to noise and clutter in the image.
Active Shape Models (ASMs) [22] are an example of deformable models of shape that
iteratively alter shape to fit the contour of the object of interest in a novel image. This
method uses a statistical model of shape that is derived from the covariance matrix of
a Point Distribution Model (PDM) built from existing training examples and a set of
means and covariance matrices of grey level profiles for each landmark point on the PDM.
The mean and principal components of the training examples are used to create plausible
examples in novel images. A model of appearance is built by sampling along the profile
normal to the contours in the training set, building a mean and covariance of the grey-level
structure at each landmark point. In novel images this is used to localize these landmark
points so that model parameters can be found that best fit the new found points. This
model of appearance assumes the texture at each landmark point is thus consistent.
ASMs have been applied to a wide range of medical applications [20, 21]. Such applications include the related field of location of vertebrae in DXA with a view to measuring
vertebral deformities caused by osteoporosis [88]. van Ginneken et al. [94, 95] used ASMs
to interpret chest radiographs using feature selection and a texture classifier to locate
points as the texture at each landmark was not consistent and edge strength was often
weak. van Ginnekens technique has also been used to segment the epicardium of the human heart in MRI images by Ordas et al. [70]. These works again highlight the difficulty
in localizing the contours of organic structures such as bone in radiographs. This was
again confirmed by Redfield et al. [74], who found the most error came from localizing
the pelvis. Redfield et al. did not apply their ASM to measuring acetabular wear.
41
As noted above ASMs require a training set of images to be annotated with landmark
points to build the PDM. Unfortunately in the case of the problem discussed in this thesis
difficulty arises in choosing sufficient landmarks points on the contour of the femoral head
and the acetabular rim. As shown previously the medial, lateral and central points of
the elliptical portion of the femoral head are obvious choices. Additionally the medial,
lateral and central points of each half of the acetabular rim could be used but this makes
only nine landmark points. As noted in Chapter 3 it is desirable to avoid using landmark
points on the pelvis to additionally constrain the model, so the choice of landmark points
appears limited.
4.3
Shape
The method described in this work uses a similar grey level model to ASMs, but fits
an ellipse directly to points located using a model of appearance. Both structures are
elliptical as the spherical femoral head projects onto the radiograph as an ellipse where
the major axis is typically 0.2mm longer than the minor axis and the acetabular rim is a
circle translated and rotated away from the center beam. Therefore ellipses are used to
localize both contours. For a more detailed treatment of ellipses and their properties the
reader is directed to Appendix A.
4.3.1
Fitting Ellipses
Having now defined the parameters and coefficients of an ellipse it is necessary to consider
methods of fitting an ellipse to a set of points. With only five points this is trivial, as an
exact conic can be computed, although this may not necessarily be an ellipse. However
when six or more points are present the task becomes increasing complex.
Fitting a geometric primitive such as an ellipse to multiple points is a common task
in computer vision. Ellipse fitting methods can be divided into two broad categories,
optimization algorithms and voting/clustering algorithms. The former are based on the
optimization of an objective function that characterizes the goodness of fit of a particular
geometric primitive. Voting/clustering algorithms can have a higher tolerance to outlying points but are always iterative and thus more computationally expensive. However
42
voting/clustering algorithms are discussed in greater detail in Chapter 5. In the case of

building the model described in this chapter it can be assumed that the user can assess
the quality of any data points they have generated and thus remove outliers.
All ellipse fitting methods considered in this chapter assume all N datapoints supplied
belong to one ellipse. The method of least squares (LS) assumes that the best-fit curve
of a given type is the curve that has the minimal sum of the squared residual errors from
a given set of data. The least squares fitting of a conic involves the minimization of an
objective, or error of fit function. One such function is shown in Equation 4.1.
a = min
N
X
Ax2 + By 2 + Cx + Dy + Exy + F
2
(4.1)
i=1
This is a commonly known error of fit function known as the algebraic distance. The
least squares minimization of the algebraic distance (LSA) and other error of fit functions
worthy of consideration are discussed in more detail in Chapter 5. It should be noted that
Equation 4.1 alone can produce a fitting that is not an ellipse but rather another type of
conic that better minimizes the error of fit. It is necessary to place a constraint for ellipsespecificity on the fit, as per B 2 4AC < 0. It must also be constrained so that it does not
return the trivial solution of a = 0. Such constraints typically require the coefficients to
be normalized, such as for example Booksteins constraint, A2 + B 2 + C 2 = 1. Fitzgibbon
and Fisher give a good review of constraints [33] that have been investigated, but it should
be noted that constraints can make the fitting procedure iterative, such as in Gander et
al. [37], where the constraint must be checked on each iteration.
Previous non-iterative methods such as those proposed by Bookstein [10] or Taubin [90]
simply fitted a general conic and thus were not constrained to be ellipse-specific. It was
suggested that to gain ellipse-specificity one should simply reject fits where the discriminant B 2 4AC indicated the conic was not an ellipse. Such methods did not perform
well in the presence of data points that better fitted alternative conics, sometimes being
unable to produce a valid elliptical solution. Fitzgibbon et al. [34, 73] proposed a noniterative method that can be solved in a closed form and guaranteed an elliptical solution.
Halir [42] showed that some of the assumptions made in Fitzgibbons work were incorrect,
most notably when all the data points lay exactly on an ellipse. Halir was able to rectify
43
these problems without sacrificing ellipse specificity or the closed form solution, and thus
this method is used throughout this work whenever an LSA fit is required. Figure 4.5(a)
shows a series of datapoints that lie almost exactly on an ellipse - thus showing why Halirs
method is preferable in this instance.
Fitting using the algebraic distance has several disadvantages. Firstly it is not invariant
under Euclidean transformation. When used in LSA fitting a simple translation of the
points alters not only the centre points but also the axes and orientation. Secondly the
algebraic distance exhibits a high degree of curvature bias, meaning that a point at a
high curvature portion of the contour contributes less to the fitting than a point having
the same amount of noise but at a low curvature portion of a contour, leading to overly
eccentric or shrunk ellipses. Thirdly it weights points on the exterior of the ellipse higher
than those in the interior.
A better fit often considered in ellipse fitting literature is the minimization of the
orthogonal distance (LSO). This is the distance between a datapoint and the closest point
on the elliptical contour, known as the orthogonal contacting point. This orthogonal
contacting point can be found by solving a quartic expression or iteratively solving the
conditions of orthogonality. The latter method is discussed in Chapter 5 where it is used
as an error of fit function in robust ellipse fitting. What is relevant in the discussion here
is that there is no closed-form solution to minimizing the orthogonal distance so one must
use a nonlinear least square algorithm, such as the generalized Gauss-Newton method.
Gander et al. [37] extended their iterative method for minimizing the algebraic distance
to the orthogonal distance. This method used a large number of fitting parameters, typically n+5 unknowns found in 2n equations from n measurement points. This was noted to
cause unnecessarily deteriorative performance of convergence as n increased. Additionally
Ganders method was not able to cope when initialised with a circle fit.
Ahn et al. [1, 2] created a method for least-squares orthogonal (LSO) distance fitting
of circles, spheres, ellipses, hyperbolae and parabolas which could be constrained to be
specific to the desired geometric primitive. This used Gauss-Newton minimization to solve
the non-linear least squares problem, requiring an initial parameter vector and step size
. An algebraic fit could provide a good initialisation but unlike the work of Gander et al.
the method was able to converge on the correct ellipse parameters given a least-squares
44
estimate of a circle as its initial parameters. The advantage of Ahn et al.s method over
other LSO methods such as that of Gander et al. is that it converges on the correct solution in fewer iterations and computes orthogonal contacting points faster than Ganders
algorithm.
Within the context of building the model of shape for the application described in this
thesis an LSO fit with an LSA initialization was used. This was mainly due to the low
curvature bias of LSO. LSA on its own was shown to hamper fits to the acetabular rim
as the curvature bias perturb fits to noisy and incomplete datasets. An example of this is
shown in Figure 4.1 where the LSA fit has generated a wildly different fit than the LSO
fit. The LSO fit better represents the dataset.
On its own an LSA fit is often adequate to fit to points placed on the femoral head,
which is a near-circular ellipse. However in the case of the acetabular rim, which is
often highly eccentric, it is desirable to use an LSO fit initialized with an LSA fit to
produce accurate results. This became readily apparent when fitting to noisy points on
the acetabular rim. Thus all ellipse fitting in this chapter used LSO ( = 1.2) initialized
by LSA.
Figure 4.1: Comparison of two fits to the same points using LSO initialised from a circle
fit and a standard LSA fit
45
4.3.2
Modelling Ellipses on Radiographs
Using the algorithms described above to fit ellipses it was necessary to supply data points
required to generate the ellipse fits. This, as noted, requires a model of appearance which
must be built from a set of training radiographs. First however the elliptical contours
must be localized on the training radiographs so that the appearance around the contours
can be modelled.
On each image in the training set a series of points was placed interactively by a
human user on the contour of the femoral head, equi-spaced and beginning at the start
of the elliptical section (as shown in Figure 4.2), and finishing at the end of the elliptical
section. An LSO fit was applied to the data points to obtain the ellipse parameters,
using the method described above. The resultant ellipse was placed around the contour
of the femoral head and displayed to the user. A second set of points was then placed
interactively on the outside contour of the acetabular rim as shown in Figure 4.3. In the
experiments detailed here qr points are used on the acetabular rim.
What resulted from this protocol was a set of ellipse parameters for the femoral head
and a separate set for the acetabular rim. a
, b, were computed separately for the femoral
head and acetabular rim ellipses for use in initialization on novel images.
4.4
Appearance
Unlike some of the applications using ASMs discussed in Section 4.2 the texture on both
the elliptical contours is consistent when normalized for intensity. As previously mentioned
there are few landmark points upon the contours. Given the apparent similarity of the
edges at all points on the femoral head and acetabular rim it was assumed that only
one model of appearance need be computed for the femoral head and only one for the
acetabular rim. Beyond this departure the method for building the model is as per the
standard method of modelling the local structure for an ASM. A brief review of this
method follows.
Once a set of ellipse parameters for a particular contour has been computed for each
radiograph in the training set N normals are extended from each ellipse. For every normal
i in all of the T images each normal is k pixels long. There are therefore T N profiles in
46
total across the entire dataset.

The first order derivative profile is obtained by differentiating a profile of grey level
values, t, where ti is the ith grey-level pixel value. This is done as follows
gi = [ti2 ti1 ... tik tik1 ]T
(4.2)
The normalized first order derivative profile is then obtained
pi = P
gi
|gij |
(4.3)
The mean of all the profiles in the images is then obtained by the following calculation
p
=
N
1 X
pi
N i=1
(4.4)
The covariance matrix of all the profiles is computed
C=
N
1 X
(pi p
)(pi p
)T
N i=1
(4.5)
The underlying assumption of this model is that the samples are distributed as a multivariate Gaussian and thus the mean and covariance are used as a model of appearance.
(a)
(b)
Figure 4.2: (a) Set of data points on the femoral head. (b) An LSO fit with 30 normals
extended from the contour (10 more were discounted)
47
This method is used on all the ellipses in the training set corresponding to the femoral
head, with normals of length k1 and then separately on all the ellipses corresponding to
the acetabular rim in the training set of length k2 . In the case of the femoral head it is
necessary to discount normals on the neck of the prosthesis and in the case of acetabular rim ellipses normals that pass through the femoral head ellipse are discarded as the
femoral head obviously occludes the rim marker.
(a)
(b)
Figure 4.3: (a) Set of data points on the acetabular rim. (b) An LSO fit with 40 normals
extended from the contour
48
Thus a mean and covariance for the femoral head and a mean and covariance for the
acetabular rim are computed. This process resulted in two distinct models of appearance.
The means and standard deviations of both these models obtained from the covariance
matrix can be seen in Figure 4.4.
These models are used to localize similar points on novel radiographs which through
(a)
(b)
Figure 4.4: Plot of the mean profiles, with error bars showing 1 SD, obtained from 50
postoperative radiographs for (a) the femoral head and (b) the acetabular rim, where k1
= k2 = 31
49
ellipse fitting would constrain the ellipse parameters to be, ideally, that of the structures
present on the radiograph. To this end it was necessary to provide an initial estimate of
the ellipse parameters, a method for adjusting these parameters and prior knowledge of
the elliptical structures to aid in locating them. Some prior knowledge was available in
the form of mean parameters obtained from the training set.
Pixel values along N search normals of length m (where m > k) are sampled. There
are m k possible positions along each normal. For each position on a normal the surrounding of k pixels are sampled. The normalized first order derivative of these pixels is
then computed as shown above. The Mahalanobis distance in Equation 4.6, a distance
measure based on correlations between a novel profile and the set of values measured from
the training set, between this observed profile, j, is calculated. Minimizing the Mahalanobis distance is akin to maximising the probability that the observed profile is from the
distribution of profiles learned during training
)T C1 (pj p
)
(pj p
(a)
(4.6)
(b)
Figure 4.5: Location of datapoints on (a) the femoral head using the mean and covariance
of the femoral head profiles, where m1 = 111 and (b) the acetabular rim using the mean
and covariance of the acetabular rim profiles where m2 = 151.
50
Thus the point with the lowest Mahalanobis distance is retained as it is the best match
to the model. Examples of this can be seen in Figure 4.4. By an ellipse fitting to these
points it should be possible for the ellipse to localize the desired contour, assuming a good
initialization is provided. Thus an interative method for localizing an elliptical contour,
described in Figure 4.6 can be proposed, with convergence criteria for when either the
change in ellipse parameters drops below a predefined threshold or a fixed number of
iterations have occurred.
In the case of the femoral head it is necessary to automatically discount datapoints
that lie in the neck region. Therefore, the derivatives are summed for each matching profile
and the 25% with the lowest sums are discarded. This number was chosen because the
neck of the prosthesis can occlude up to approximately one-quarter of the femoral heads
contour.
In the case of the acetabular rim points are discarded that are within the femoral
head. This is done by calculating the Euclidean distance from the femoral head ellipses
centre to the datapoint and to its orthogonal contacting point on the femoral head. If the
contacting point is closer to the ellipse centre then the datapoint, then the datapoint is
outside the ellipse. If on the other hand the contacting point is further from the ellipse
centre the datapoint, or the contacting point and datapoint are identical, the datapoint is
within the femoral head and should be excluded.
4.5
Summary
This chapter describes a method for modelling the shape and appearance of the contours
of the femoral head and acetabular rim. The method is akin to that of an ASM, but
1. Obtain initial estimate of ellipse parameters.
2. Extend normals to the ellipse to find the best matches for the model of
appearance.
3. Update the parameters (a, b, xc , yc , ) to fit the located points.
4. Repeat 2-3 until convergence criteria fulfilled.
Figure 4.6: The Active Ellipse Algorithm
51
due to the lack of good landmark points upon the contours an alternative approach was
necessitated. Rather than build a PDM the shape of each elliptical contour is modelled
parametrically by taking the mean axes and orientation of T ellipses. These ellipses are
obtained by LSO fits to points manually placed on the contours of the structures.
The algorithm for locating the elliptical contours in novel images is summarised in
Figure 4.6. A series of points are computed on the ellipse from which normals are extended.
The normalized first-order derivative profile is calculated along each normal. In the case of
the femoral head it is necessary to discount normals on the neck of the prosthesis, while in
the case of the acetabular rim it is necessary to discount normals where the acetabular rim
is occluded by the femoral head. A covariance matrix and a mean profile are determined
from all these normalized profiles. This results in two distinct models, one for the femoral
head and one for the acetabular rim.
Chapter 5 details the process of gaining an initial estimate of xc and yc , the centre of
the femoral head ellipse, as these cannot be determined from the training set. Once this
is located the model should converge on the femoral heads contour by obtaining points
by measuring the Mahalanobis between each possible profile on the search normal and
the model of appearance. Fitting to these datapoints using the LSO and LSA methods
described in this chapter is undesirable, and the reasons for this are also described in
Chapter 5, along with an alternative solution.
52
Chapter 5
Parameter Estimation
In Chapter 4 a model of appearance of the desired contours was created. The mean ellipse
parameters, a
, b, , were learned for each of the contours. However the centre points for
the ellipses, xc and yc were left undefined as the mean of these parameters would not prove
useful in initialising searches.
This chapter details a method for automatically determining initial estimates of the
centre points for the femoral head to initialize the corresponding active ellipse. As the
method proposed in Figure 4.6 is a local search a good initialization is required. The
intention is to locate the femoral head and then the acetabular rim serially. To initialize
the search only the centre points of the femoral head are initially required. This chapter
also details the process of estimating more accurate elliptical parameters for the femoral
head and acetabular rim contours on novel radiographs. In doing so it examines the need
for robust estimation of the ellipse parameters for the femoral head and acetabular rim.
Estimation of the ellipse parameters is closely related to statistical techniques of regression. The aim of ellipse fitting, as with most model fitting, is to find a model which
best agrees with the given data set. To measure agreement between the ellipse and the
data set a residual measure is required. This is a distance measure between the model and
the data points, or rather a measure of the error of fit. There is therefore typically one
or more residual value for each data point in the data set. The ideal parameters should
minimize a function of the error of fit. In traditional statistics the residual is typically the
orthogonal distance between the model and the data point. In the case of estimating ellip-
53
tical parameters this is not common due to the complexity of computing this orthogonal
distance.
This chapter introduces ellipse fitting techniques that are available both from statistics
and computer vision literature. Each technique has a breakdown point, a percentage
of outlying points that can be present within the data set before arbitrarily incorrect
fits are returned. LS estimation is examined and compared with a method for robust
parameter estimation, Least Median of Squares (LMedS). The difficulties of applying LS
and LMedS to ellipse fitting is highlighted by explaining the need for different error of fit
(EOF) functions, and by a direct comparison with several EOF functions, including the
orthogonal distance.
5.1
5.1.1
Initialization
Template Matching
In template matching techniques a template image is transformed relative to the target

image. A distance function is applied to each possible transformation to measure the
dissimilarity of the template and the image. The location with the smallest dissimilarity
is retained. Numerous functions exist for this, such as a simple Euclidean distance function,
the sum of the absolute differences, the sum of the squared differences or the normalized
cross correlation.
P P
i
C(x, y) = rP P
i
Ix+i,y+i Ti,j
I2x+i,y+j
P P
i
T2i,j
(5.1)
where C is the correlation between the template and the current translation, I is the target
image, T is the template image and x,y are the co-ordinates of the current translation on
the target image.
In Equation 5.1 the correlation between template and image is normalized and thus
any local maxima defines solely regions of similarity. The normalized cross-correlation was
chosen rather than the correlation or the Euclidean distance as these converge on areas
where the image values are high, regardless of any similarity to the template. In the case
of locating the femoral head this can cause the template to locate the edges of the film.
54
Figure 5.1: Template image and subsampled template
Figure 5.2: Successful initialization by template and resultant convergence
55
(a)
(b)
Figure 5.3: Region of interest and full view of radiograph where failure to localise femoral
head contour is due to quantisation effect caused during scanning. This manifests itself
as a streaking effect across the entire radiograph.
(a)
(b)
Figure 5.4: (a) Region of interest of two radiographs with noise shown after histogram
equalisation and (b) quantisation effect caused during scanning. This prevents localization
of the contours.
56
(a)
(b)
Figure 5.5: Region of interest of two radiographs with poor levels of contrast. (a) This
radiograph was not scanned on a backlit scanner, resulting in the dark image and (b)
the femoral head is almost occluded by the surrounding bone. The template is unable to
locate the femoral head and instead favours the symphisis pubis.
(a)
(b)
Figure 5.6: Region of interest of two radiographs with poor levels of contrast. (a) Poor
contrast prevents accurate localisation of the femoral head. (b) a poorly contrasted radiograph containing a metal backed acetabular cup.
57
As a template image a single total hip replacement radiograph cropped to 200 x 200
pixel image and centred precisely on the femoral head was used. One of the disadvantages
of template matching is it is slow when a large template is convolved with an even larger
image. Therefore the template and target image are both subsampled to produce a coarse
image and template. On a 150 dpi image the template and target image are both scaled
to 10% of the original size.
The location with the highest similarity to the template is used as the initial centre
points (xc , yc ) for the femoral head ellipse with a, b and set to the mean values obtained
during training.
The disadvantage of using template matching is that it in theory requires a single
template for each prosthesis model. Automated initialization removes any intra-personal
and inter-personal variation in measuring wear using the active ellipse method.
5.1.2
Results
Evaluation of the template matching was conducted on two data sets, both scanned at 150
dpi and containing Zimmer CPT prostheses. The first dataset of 100 images was cropped
to include a 600 by 600 region of interest and contained femoral heads of 28mm. The
second dataset consisted of 177 uncropped AP radiographs containing the entire femur,
prosthesis and in some cases radiopaque boundaries of the plate, with an unknown femoral
head diameter.
A successful initialization using the template was defined as convergence around the
femoral head from the initial parameters using the initialisation p. On the cropped dataset
Table 5.1 shows 100% success, while on the uncropped dataset seventeen failed. Ten successfully segmented the femoral head when given user-defined initial centre points within
the femoral head. The remaining seven were still left unsegmented. Three failed due
to noise introduced when scanning the film creating false edges throughout the image.
These are shown in Figure 5.3-5.4. Three failed due to there being no contrast between
the femoral head contour and the surrounding tissue, while one failed as the radiograph
contained a metal backed acetabular cup. These are shown in Figures 5.5-5.6
58
Dataset
Success
Failure
600x600
100
Full Film
160
17
Table 5.1: Template Matching Success
5.2
Data Points and Outliers
In Chapter 4 it was explained how the model of appearance was used to localize points on
the desired contours. However as can be seen in Figure 5.7(a) not all points found in the
search correspond to points on the contour of the structure. In Figure 5.7(a) outliers are
prevalent due to (1) radiopaque clutter, in this case a pair of screws to aid fixation of the
acetabular cup, (2) points located on the contour of the neck of the prosthesis (typically
among the 25% of points automatically discarded) and (3) points located within the neck
of the prosthesis, where the contour of the femoral head is occluded by the neck and
thus there is no edge information. With no corresponding contour available the model
instead matches to noise within the neck that has been amplified by the normalization
process. These are typically discarded by the 25% rule but shown here for clarity. Lastly
erroneous points can be found where (4) the contour of the femoral head is occluded by
the acetabular rim and thus points are placed on the contour of the rim.
Outliers are a serious concern as the LS fit has a breakdown point of 0% outliers. Figure
(a)
(b)
Figure 5.7: (a) Set of data points on the femoral head. (b) The resultant LS fit
59
5.8 shows a dramatic example of this where an LS ellipse fit is peturbed when there is but
one outlying point and all remaining points are exactly on the original ellipse contour. LS
assumes that the noise corrupting the data is Gaussian with zero mean. This is desirable
for computational efficiency as the regression parameters can be solved analytically, but
at the expense of the method being highly sensitive to non-Gaussian noise with even a
single outlying point resulting in a poor solution.
The datapoints returned by the models built in Chapter 4 unfortunately contain outliers and thus the LS fits are perturbed. What is required is a method that is robust,
capable of tolerating potentially large amounts of outliers, or to quote Huber [49] displaying insensitivity to small deviations from the assumptions of LS fitting.
5.3
Robust Parameter Estimation
Rosin [78] proposed a method for robustly estimating ellipse parameters by five point fits,
the minimal subset of the set of datapoints required to estimate ellipse parameters. By
obtaining a fixed number of hypotheses and retaining their ellipse parameters Rosin chose
the median of each parameter as a final estimate. However this method was later noted
by Rosin to have a breakdown point of approximately 12.9% outliers. Rosin additionally
noted this was inadequate and statistically inefficient, particularly in cases such as when
the ellipse parameters are near-circular and thus was meaningless. Additionally all 5
parameters are estimated separately, meaning there was little or no correlation between
(a)
Figure 5.8: A single outlying point distorts the LS fit (dashed line) from the original ellipse
60
parameters.
The Hough Transform method incorporates a degree of robustness but is optimal
at estimating parameters when there are three or less unknown parameters. Roth and
Levine [82] applied the Least Median of Squares (LMedS) algorithm to ellipse fitting.
Both Rosin [80] and Zhang [102] indicated that that in the presence of outliers the LMedS
approach produced the best results when compared to other ellipse fitting methods including LS, Rosins previous median method [78]. Rogers and Graham [76, 77] used LMedS for
robust active shape model searching, concluding that it, with RANSAC and the Weighted
Least Squares Huber M-Estimator as alternatives, improved the accuracy of the search.
LMedS can be summarized by the following
min median EOF2
(5.2)
where EOF can be any measure of the distance of a datapoint from the elliptical contour.
In Chapter 4 the EOF function proposed was the algebraic distance. A more detailed
description of the LMedS algorithm is given in Figure 5.9.
The fit to the minimal subset that LMedS provides can be improved upon. The noise of
the dataset is estimated by using the median absolute deviation (mad) to the differences
of the errors from the median of the errors
mad = median |EOFi median EOF|

1. Sample minimal set of features by choosing random subsamples from the
dataset.
2. Instantiate model hypothesis.
3. Calculate squares of the errors of fit.
4. Retain median of squares of the errors of fit and the associated model hypothesis.
5. Repeat until maximum number of hypotheses (h) obtained.
6. Choose hypothesis with lowest median square of the errors of fit as an estimate.
Figure 5.9: The LMedS Algorithm
61
(5.3)
This is modified by a Gaussian normalisation and finite-sample correction factor

f = 1.4826 1 +
5
n1
(5.4)
Points outwith a given threshold of 3f mad are weighted w = 0, where w is the weighting
of the points such that
w=
|EOFi | < 3 f mad
otherwise
(5.5)
This is a non-linear minimization problem and thus must be solved iteratively. Therefore as a statistical estimator LMedS is considerably less computationally efficient than
LS, but it has a breakdown point of 50%, considerably higher than LS.
The number of hypotheses, h must be chosen. The following equation assumes that
the datapoints consist of a fraction of outlying data points, , and that the probability
(P ) of randomly selecting p points that are all inliers is
P = 1 (1 (1 )p )h
(5.6)
Therefore rearranging Equation 5.6
h =
log(1 P )
log(1 (1 )p )
(5.7)
However values given by Equation 5.7 can often be optimistic. This equation gives the
minimum number of hypotheses consisting purely of inliers but does not take into account
the quality of the inliers. Noisy inlying points require h to be much larger.
LMedS is very similar to the Random Sample Concensus (RANSAC) method of Fischler and Bolles [32]. Bolles and Fischler applied RANSAC to estimate the known parameters of elliptical contours on cylinders [9]. RANSAC and LMedS differ in their convergence
criteria. RANSAC requires a fixed number of hypotheses and a threshold which is used
to determine whether the current hypothesis is so acceptable that the algorithm should
retain this hypothesis and end the random sampling. LMedS on the other hand requires
only a fixed number of hypotheses. This becomes advantageous when testing different
62
EOF functions as it is difficult to work out a threshold value that is equivalent for all the
EOF functions to ensure fair testing.
In implementing this algorithm for ellipse fitting, additional consideration is required
in Step 3. As noted previously the error of fit is typically measured by the orthogonal
distance between the model and the data point, but this is not trivial to compute when
estimating ellipses. To this end a considerable amount of research has been undertaken
into creating approximations of this orthogonal distance. A brief review follows outlining
the most noteworthy EOF functions. For a more comprehensive review the reader is
directed to the work of Rosin [79, 81] and Safaee-Rad et al. [83].
5.4
Review of Ellipse Error Functions
Choosing the error of fit function to use in any ellipse fitting is important as a poor
choice can result in biased solutions. Examples of biases present in existing EOFs include
curvature bias and asymmetry. Curvature bias in an error fitting algorithm affects the
weight of points near the ends of the ellipse in an ellipse fit. Curvature bias can adversely
affect fits to eccentric ellipses, such as those of the acetabular rim. Asymmetry affects the
weighting of points inside the ellipse compared to those outside the ellipse.
Rosin [79, 81] suggested 13 error of fit functions and measured the curvature bias and
asymmetry of each one. Additionally linearity was measured. An EOF function with a
linear relationship between its values and the actual Euclidean distance values would have
a high linearity. Visualisations were obtained by comparison to a distance value thought
to be close to the orthogonal distance that was obtained by plotting the ellipse contour
into an image and measuring the Euclidean distance from the closest point on the pixelized
ellipse to the desired point. EOF3 , the angular bisector of the foci, was found to be the
best of the studied EOF. Of the 13 error functions the angular bisector of the ellipse foci
was considered to provide the best goodness of fit, and the lowest curvature bias, with an
acceptable level of asymmetry. Safaee-Rad [83] similarly compared 4 error of fit functions
to the orthogonal distance. The orthogonal distance is usually taken as the gold standard
by which other EOF functions are compared. The experiments are therefore restricted to
the other EOF functions and no qualitative evidence exists as to the orthogonal distances
63
performance as an EOF in robust ellipse parameter estimation.

Four EOF functions are reviewed and evaluated in this chapter. The algebraic distance
and weighted algebraic distance are the most commonly used, while Rosin showed the foci
bisector distance to be the best EOF. The use of the geometric distance is believed to be
novel to this thesis.
5.4.1
Algebraic Distance
The algebraic distance of a point (xi , yi ) from an ellipse contour is
EOF1 = Ax2i + Bxi yi + Cyi2 + Dxi + Eyi + F
(5.8)
The primary advantage of using the algebraic distance is computational efficiency. As

noted in Chapter 4 the algebraic distance has a closed form LS solution whereas all other
error functions must be minimized iteratively.
5.4.2
Weighted Algebraic Distance
The weighted algebraic distance is simply the algebraic distance inversely weighted by its
gradient. The gradient is calculated using the first order partial derivatives as follows
x
2Axi + Byi + D
=
y
2Cyi + Bxi + E
(5.9)
Therefore this is
EOF2 =
Ax2i + Bxi yi + Cyi2 + Dxi + Eyi + F

x
y
(5.10)
While this has been shown to exhibit a lower curvature bias than EOF1 it should be
noted that EOF2 exhibited higher asymmetry in Rosins evaluations. Points within the
ellipse are weighted more than points outside the ellipse.
5.4.3
Angular Bisector of the Foci
The foci are a pair of fixed points inside the ellipse that can be used in defining the curve
of the ellipse. When considering a locus of points on the contour of the ellipse the sums
of the distances to both foci are constant for all the points.
64
In order to determine the foci of an ellipse the following equation is used
f1 =
q
cos
(a2 b2 ), 0
sin
sin cos
f2 =
q
(5.11)
cos
(a2 b2 ), 0
+ [xc , yc ]
sin
sin cos
+ [xc , yc ]
(5.12)
Rather than calculate the orthogonal contacting point of the datapoint the contacting
point of the angular bisector of the datapoint to each of the foci, f1 and f2 is used instead.
The closest point of intersection, xf , yf , is found and the distance between it and xi is
taken as an error of fit.
EOF3 =
5.4.4
(xf xi )2 + (yf yi )2
(5.13)
Orthogonal Distance
The orthogonal distance between a datapoint and an ellipse is the smallest Euclidean
distance among all the distances between the point and the ellipse. This is typically the
EOF function of choice in applications such as line estimation or circle fitting. In ellipse
fitting however the difficulty in implementing such an EOF function lies in locating the
orthogonal contacting point.
Both Safaee-Rad et al. [83] and Ahn et al. [1, 2] proposed methods for locating the
orthogonal contacting point. In both methods an xy coordinate system is defined where
Figure 5.10: Location of the closest point of intersection of the angular bisector (xf )
65
the major axis lies along the x axis and the minor axis along the y with the ellipse centred
at the origin. Thus the parameters xc , yc and are removed. The ellipse is solely described
by its axis lengths a and b and the datapoints are translated and rotated accordingly. The
equation of the ellipse is now
x2 y 2
+ 2 =1
a2
b
(5.14)
The product of the gradients of two orthogonal lines is always -1. Thus the following
relationship should be true between the data point (xi , yi ) and its orthogonal contacting
point (x,y)
y yi y
b2 x yi y
= 2
= 1
x xi x
a y xi x
(5.15)
The orthogonal contacting point satisfies the condition in Equation 5.14, ensuring that
the point is on the elliptical contour, and that of Equation 5.15. Equations 5.14 and 5.15
(a)
(b)
Figure 5.11: Initialisation (x0 ) and successful location of the orthogonal contacting point
(x)
66
can be rearranged respectively as follows

1
f1 (x, y) = (a2 y 2 + b2 x2 a2 b2 ) = 0
2
(5.16)
f2 (x, y) = b2 x(yi y) a2 y(xi x) = 0
(5.17)
Both these conditions must be true for the orthogonal contacting point. Safaee-Rad [83]
et al. combined the two equations into one quartic equation. This equation could provide
4 solutions and the solution with the shortest distance to the datapoint was taken as the
orthogonal contacting point. However this method was unstable when the datapoints were
close to the axes and instead Ahn et al. used the generalized Newton method to minimize
the derivatives of the two orthogonal contacting point conditions
f1 f1
x y
Q=
f2 f2
x y
(5.18)
Therefore
Q=
b2 x
a2 y
(a2 b2 )y + b2 yi
(a2 b2 )x a2 xi
(5.19)
Qk x = f (xk )
(5.20)
xk+1 = xk + x
(5.21)
An initial value is required for this process, given by the standard point of the ellipse.
This is shown in Figure 5.11 as the point where a line between the centre of the ellipse
and the data point intersects the elliptical contour. This can be found as follows
xi
ab
yi
x0 = q
b2 x2i + a2 yi2
(5.22)
This is a good initialisation as it always lies in the same quadrant of the ellipse as the
orthogonal contacting point. Ahn et al. noted that after 3-4 iterations convergence was
67
reached upon the orthogonal contacting point. Thus this method is less computationally
expensive than solving Safaee-Rads quartic equation.
It should be noted that Q will be singular when a = b and xi , yi lies on the centre of
the ellipse. In this instance there is no one orthogonal contacting point and any point on
the circle is a valid contacting point.
With the orthogonal contacting point located the Euclidean distance is given as
EOF4 =
(x xi )2 + (y yi )2
(5.23)
Figure 5.11 shows a sample ellipse with orthogonal contacting point successfully calculated for the datapoint.
5.5
Evaluation
The work of Rosin provides a good indication of the expected characteristics of all the
discussed EOF functions save the orthogonal distance. However there appears to be no
previous work using the orthogonal distance as an error function for LMedS ellipse fitting.
Therefore a quantitative test was required. Each of the above EOF was tested with
synthetic data to demonstrate each EOF functions effect on the accuracy and robustness
of the LMedS algorithm.
5.5.1
Synthetic Data
Synthetic noisy data sets were created from known ellipse parameters as per prior experiments conducted by Rosin [80]. Unlike in Rosins study two ellipses were considered with
eccentricities of 0.66 and 0.98 (see Figure 5.12). Three types of data set were created from
each of the two ellipses.
Gaussian noise Data points were sampled at uniform intervals along the ellipse with
additive Gaussian noise in the direction normal to the ellipse contour. Data sets
were created with = 0, 10, 20, 30, 40, 50, 60 and 70.
Gaussian outliers Some points were sampled with noise drawn from a low variance
Gaussian ( = 5) while the remainder had high variance Gaussian noise ( = 500)
68
(a)
(b)
Figure 5.12: The two ellipses used to generate synthetic data. Both were centred at the
origin and aligned with the image axes. (a) The less eccentric ellipse, a = 333, b = 250.
(b) The more eccentric ellipse, a = 300, b = 60.
thus creating outlying points. Data sets were created with the percentage of points
with high variance noise set to 0%, 10%, . . ., 90%.
Structured outliers This type of data set was designed to simulate structured noise by
sampling some of the points from straight line segments close to the ellipse. All
points were sampled with Gaussian noise ( = 5) from a closed contour, 80% from
an elliptical arc, 10% from a line segment orthogonal to that arc, and 10% from
another line segment rotated 45 with respect to the first line segment. Data sets
were created with the percentage of structured outliers set to 0%, 10%, . . ., 90%.
An example from each type of data set is shown in Figure 5.13. Each example consisted
of 38 points. LS fits using orthogonal and algebraic error functions and LMedS fits (with
LS fine tuning on resulting inliers) using algebraic, weighted algebraic by gradient, foci
bisector distance Ros02 and orthogonal error [2] functions were performed. The Euclidean
distances between the original and recovered centre points were used as a measure of
accuracy.
5.5.2
Radiographic Data
The elliptical projections of acetabular rims in standard clinical radiographs typically have
eccentricities between 0.8 and 1.0. When the eccentricity is 1.0 the circular contour of the
acetabular rim projects as a line and the method cannot be used to localise the contour.
69
(a)
(b)
(c)
Figure 5.13: Data points created from the more eccentric ellipse. (a) Gaussian noise
( = 10). (b) Half low variance Gaussian noise ( = 5) and half high variance Gaussian
noise ( = 20 in this visualisation). (c) Structured outliers sampled from noisy line
segments
Figure 5.14: An acetabular rim with eccentricity 1.0. The method cannot localise such a
contour.
An example of this is shown in Figure 5.14.
A set of 19 radiographs containing Zimmer CPT prostheses with particularly eccentric
rim projections (> 0.96 and < 1.0) was obtained. The most accurate error-of-fit functions
from the experiments using the synthetic data sets described above were selected. These
were used to perform active ellipse localisation on the radiographs. Localisation was run
twice on each image, providing 38 results per method. No LS fine tuning was performed
on the resulting inliers.
70
5.6
5.6.1
Results
Synthetic Data
Figures 5.15-5.17 show the alpha trimmed means ( = 0.1) of the centre errors for each
of the synthetic data sets. Each point on these plots was computed from 500 examples.
Figures 5.18-5.20 show the standard deviation of the 500 centre errors obtained per dataset.
In Figure 5.17 the centre error goes beyond the scale of the graph as selecting a large
proportion of points from the clutter resulted in extremely eccentric and erroneous ellipses
being generated. The standard deviations of the LMedS fits on the structure noise increase
around the 40%-60% margin before tailing off as the methods begin fitting to points purely
on the noisy line segments, while those of the Gaussian outliers remain very high due to
the erratic nature of the noise.
Unsurprisingly and for the reasons noted above in the presence of pure Gaussian noise,
LS outperformed LMedS irrespective of the EOF function used (see Figure 5.15). LS using
the orthogonal error performed the best, but LS Algebraic is very close behind it. On the
more eccentric ellipse the difference between LSA and LSO became more pronounced, with
LSO performing better. The LMedS orthogonal and LMedS foci bisector methods were
least accurate in this case.
However LMedS fitting was more accurate in the case of outliers, whether Gaussian
or structured. The plots in Figures 5.16 and 5.17 show that LMedS using foci bisector
or orthogonal EOF functions performed best on both ellipse eccentricities. The only
noticeable difference between these two methods before the breakdown point was at 40%
structured outliers in Figure 5.17(b). However by looking at Figures 5.21 and 5.22 it
appears that the orthogonal distance outperforms the foci bisector slightly.
Results obtained with greater than 50% outliers lie beyond the theoretical break-down
point of LMedS and so not surprisingly are poor. LMedS fitting with algebraic and or
weighted algebraic functions performed very poorly in the presence of structured noise (see
Figure 5.17). In fact LS methods were better in this case as LS fitted to both the straight
line segments and the elliptical arc, evidenced by the relatively low standard deviations of
both LS fits. The LMedS algebraic and weighted algebraic tended to favour points on the
line segments and thus had very high standard deviation.
71
(a)
(b)
Figure 5.15: Centre errors for (a) less eccentric and (b) more eccentric synthetic ellipse
data as a function of
72
(a)
(b)
data with Gaussian outliers
73
(a)
(b)
data with structured outliers
74
(a)
(b)
Figure 5.18: Standard deviation of centre errors for (a) less eccentric and (b) more eccentric
synthetic ellipse data as a function of
75
(a)
(b)
synthetic ellipse data with Gaussian outliers
76
(a)
(b)
synthetic ellipse data with structured outliers
77
(a)
(b)
data with Gaussian outliers, visualising only foci bisector and orthogonal distance error
functions
78
(a)
(b)
data with structured outliers, visualising only foci bisector and orthogonal distance error
functions
79
Figure 5.23: Failure of the method as the highly eccentric rim is at such an extreme
orientation and thus too far from the mean rim parameters
5.6.2
Radiographic Data
On the radiograph dataset, LMedS Orthogonal failed 15 times out of 38, LMedS Foci
failed 16 times while LMedS Algebraic failed 32 times. Figures 5.23-5.26 show examples
of the failures for even the orthogonal distance. Given the difficulty of the data set and
the absence of any LS fine tuning to inliers identified from the minimal subset (which
increases the performance of all three error functions), the results using LMedS geometric
fitting were encouraging. An example of the output of each of these algorithms is shown
in Figure 5.27.
5.7
Summary
This chapter has described a method for providing a rough estimate of the femoral head
centre through the use of template matching. The template matching has been shown to
be 100% successful when used on a dataset cropped to show only the region of interest
for the system. On a less constrained dataset where the femoral head diameter was not
known the success rate dropped to 90%. The majority of these failures were due to poor
80
Figure 5.24: Failure of the method as the upper end point of the ellipse is occluded by
clutter
Figure 5.25: Failure of the method as the fitting has favoured the near circular wire marker
on the periphery of the acetabular cup instead of the rim
81
Figure 5.26: Failure of the method as the fitting has favoured the near circular wire
marker on the periphery of the acetabular cup instead of the rim, as the rim wire marker
is occluded
(a)
(b)
(c)
Figure 5.27: An eccentric rim with (a) failed algebraic, (b) successful foci bisector and (c)
successful geometric fits.
82
image quality.
The model of appearance described in Chapter 4 can find false matches, outliers, that
peturb the parameters found by a standard LS ellipse fit. LMedS fitting is thought to
be the best estimator in terms of robustness and accuracy. Unlike other applications of
the LMedS algorithm additional care must be taken when choosing an EOF function.
The orthogonal distance of the datapoint to the closest point on the ellipse contour is
non-trivial to compute and in the past required solving a quartic equation.
This chapter described four EOF functions. These include the algebraic distance,
gradient weighted algebraic distance, foci bisector and an alternative method of finding the
orthogonal distance by minimizing the derivatives of the two conditions of the orthogonal
contacting point via the Generalized Newton method. The performance of the LMedS
algorithm using these EOFs was evaluated on synthetic datasets. These experiments
demonstrated that in the presence of non-Gaussian noise, LMedS orthogonal fitting tended
to perform slightly better than the foci bisector distance approximation. However, it is
a computationally expensive method and in applications where speed is important the
foci bisector distance is recommended as an EOF function. In the case of the application
of measuring acetabular wear speed is not necessary thus the orthogonal distance was
retained as an EOF function.
83
Chapter 6
Eccentricity Error and Pose

Estimation
The acetabular rim wire, unless there are deformities, can be modelled by a torus. This
torus is very thin, the diameter of a ZCA acetabular cups rim wire being just 3mm [24].
Thus the outer contour of the rim wire can be approximated as a circle in 3-D. However
translation from the beam centre and rotation in depth causes this circle to project as an
ellipse on the radiograph as shown in Figure 6.1. Figure 6.2 shows a rotation in depth
performed on a ZCA cup CAD model. The projection of a circle is that of an ellipse
with differing axes if the circle lies on a plane that is not perpendicular or parallel to the
image plane. Otherwise if it lies on a plane parallel to the image plane it projects as a
circular ellipse, and if it lies on a plane perpendicular to the image plane it should project
as a straight line segment. Huang et al. [48] proved that this is even true under a weak
perspective system.
Figure 6.1: Projecting both a circle and an ellipse from a circle
84
(a)
(b)
Figure 6.2: The acetabular cup with (a) circular rim and (b) rotated in depth to produce
an elliptical rim
The centre of the projected ellipse does not correspond, in general, to the projection of
the centre of the 3-D circle. This has serious implications for 2-D measurements of wear
using the acetabular rim ellipse centre as it means the rim reference point is not fixed,
but rather varies based on the eccentricity of the ellipse projected on the radiograph. It is
therefore necessary to find the true centre of the acetabular rim and use this as a reference
point. The distance between the centre of the projected ellipse and the projection of the
centre of the circle is referred to as eccentricity error.
6.1
Eccentricity Correction
Figure 6.3 shows the eccentricity error due to the projected points B and C. C is the
projection of the centre of the circular object, while B is the centre of the projected
ellipse. The x , y , z axes is the object co-ordinate system, while x, y, z refer to the film
coordinate system and x0 , y 0 , z 0 represent the film co-ordinate system rotated such that
the x0 , y 0 are coincident to the ellipse axes.
Ahn et al. [3] proposed a method for correcting the eccentricity error. While their
method assumes a camera model, where the film precedes the object in a perspective projection model, the geometry is also applicable to the problem at hand. Several assumptions
are made that simplify the original method by Ahn et al. to apply to the situation shown in
Figure 6.3 and thus allow automated correction to be undergone without user interaction.
85
The first assumption is that the beam centre lies on the centre of the acetabular rim
ellipse. The true beam centre lies on the symphysis pubis, which can be as much as 900
pixels away from the acetabular rim centre on a 150 dpi radiograph. Ahn et al.s original
equations for eccentricity error correction are as follows
x0 =
xi
cos()FFDOFD

cos()FFDOFD 2
FFD
y0 =
sin2
(6.1)
sin2
FFDOFD
cos()FFDOFD

cos()FFDOFD 2
FFD
sin cos
(6.2)
sin2
where xi is the offset of the object target from the beam centre, r is the radius of the cup,
is the rotation in depth and is the camera rotation, shown in Figure 6.4. By assuming
the beam centre lies on the centre of the acetabular rim ellipse xi = 0 and thus x0 = 0.
To place bounds upon the size of the eccentricity error and thus see if it was worth in-
Figure 6.3: Side view of the object and image plane, aligned with y 0 to show the projection
of the entire minor axis. This shows the projection of the minor axis b, with eccentricity
error y0 shown as the distance between projected ellipse centre B and actual projection
of the centre C. In this instance =.
86
vestigating more thoroughly several cases were tested using Equations 6.1 and 6.2. These
are summarized in Table 6.1. The values of and are the maximal values given in Ahn
et al. [3], while the upper and lower values for r are taken from the ZCA cup documentation [24] and the upper and lower case values for the FFD and OFD are taken from
Krismer et al. [59]. Values as large as 2.60mm serve to highlight the concerns over the
level of systematic error that could be being introduced into the method. These theoretical
worst case values are clearly much larger than the expected rate of wear over the lifespan
of a prosthesis.
The assumption can be made when the FFD is large relative to the objects
Figure 6.4: Visualisation in 3-D of x-ray system from focus to object in a case where the
ellipse is not at the beam centre. Finally 6= for visualisation purposes. When =
this diagram collapses down.
Cup Radius (r)
FFD
OFD
y0 (mm)
24
900
270
60
30
-1.51
31.5
900
270
60
30
-2.60
24
1000
225
60
30
-1.11
31.5
1000
225
60
30
-1.91
24
1300
180
60
30
-0.69
31.5
1300
180
60
30
-1.19
Table 6.1: Eccentricity error values for given parameters.
87
depth. The exposure distance of the target cos( )FFD-OFD. Given the assumption
this is simply FFD-OFD. It is therefore possible to approximate Equation 6.2 as
follows
y0 =
FFD sin cos

FFDOFD
r
2
(6.3)
sin2
This method could be applied to correct the eccentricity error of the acetabular rim
centre. However the pose of the circle in the radiograph, or to be more specific the rotation
in depth, , must first be obtained.
6.2
Pose Estimation
In order to find the image centre of the acetabular cup it is necessary to find the pose of
the cup in the radiograph. Methods that suggest themselves include the use of a CAD
model to estimate pose or, given that the acetabular rim wire is a circle in 3-D attempting
to recover the pose of the original circle. A brief consideration of both methods follows.
6.2.1
CAD Model Based Pose Estimation
One method for accurate estimation of the pose of the acetabular rim would be to use a 3D model of the component and overlay the model with the contour of the cup in a manner
similar to the model based RSA discussed in Chapter 3. Such models are often available
from implant manufacturers. Unfortunately the contour for cemented acetabular cups is
not readily apparent in radiographs, leaving only the radiopaque wire on the rim and the
edge of the cup. This makes it difficult to accurately fit a 3-D model to the acetabular
cup contour, while it is certainly feasible for metal backed cups.
6.2.2
Pose of a Circle
Pose can be estimated via the following method [35, 92]. This algorithm has been successfully employed in estimating the 3-D facial pose of human heads in images [101] and
in sentient spaces [26], but not to correcting eccentricity error.
The general equation of a conic is
88
Figure 6.5: Typical geometry for recovering the pose of a circle from a camera setup. Note
x0 and y 0 are aligned with the axes of the ellipse on the film.
Ax2 + Bxy + Cy 2 + Dx + Ey + F = XT EX = 0
(6.4)
where X = [x, y, 1]T , all distances are expressed in multiples of the FFD and E is the
real, symmetric matrix of the ellipse as follows
E = B/2
B/2 D/2
C
E/2
D/2 E/2
(6.5)
This matrix is used to estimate two rotations and translation. The first rotation, R1
transforms the initial coordinate system 0 x y z into 0 x0 y 0 z 0 , a co-ordinate system aligned
with the major axis and minor axis with the ellipse centred at the z 0 axis. This is shown in
Figures 6.6a and 6.6b and occurs by obtaining the eigenvalues 1 , 2 , 3 and eigenvectors
e1 , e2 , e3 of E. The first rotation matrix R1 is obtained by
R1 = [e1 |e2 |e3 ]
(6.6)
The second rotation concerns the rotation in depth of the circle to become the projected
ellipse. This can be obtained by imposing the equality of the coefficients of x2 and y 2 ,
89
(a)
(b)
Figure 6.6: (a) A rotated and translated ellipse and the resultant ellipse axes x0 , y 0 , z 0 (z 0
not pictured) obtained from using the eigenvalues of the symmetric matrix as coefficients
as outlined in Equation 6.7. When expressed in the x0 , y 0 , z 0 coordinate system the ellipse
is aligned with the axes and centred on the z 0 axis (c) The resultant axes x00 , y 00 , z 00 (z 00
not pictured) obtained by rotating in depth using in Equation 6.8
90
resulting in a rotation around the x0 axis. The ellipse was originally a circle and could be
represented by the following equation of a cone
2 x2 + 1 y 2 + 3 z 2 = 0
(6.7)
where 1 = 2 . This is not currently the case and therefore the rotation in depth is
used to equate the first two coefficients such that
2 x2 +(1 cos2 +3 sin2 )y 2 +(1 sin2 +3 cos2 )z 2 +2(1 +3 ) sin cos yz = 0 (6.8)
From equation 6.8 it must therefore be true that for this to be a circle
2 = (1 cos2 + 3 (1 cos2 ))
(6.9)
This can be rearranged to find

s
= cos1
2 3
1 3
(6.10)
The second rotation is therefore identified as follows
cos
R2 =
0 sin
1
sin 0 cos
(6.11)
However the exact value of is ambiguous, lacking constraint within the four quadrants.
Plausible values of include , , + or . For the purposes of the measurements made in this and the next chapter is constrained to be within the first quadrant.
Therefore the global rotation matrix, R, is R = R1 R2 . The normal to the circle plane is
0 R13

n = R 0 = R23
R33
However it is the rotation by the x0 axis, , that is required in Equation 6.3.
91
(6.12)
6.3
Experiment
The method described above was used on a test set consisting of 50 radiographs containing
Zimmer CPT prostheses with a 28mm head diameter with the acetabular rim ellipse
already localized. Each radiograph was digitised at 150 dpi and each pixel on the digitised
radiographs corresponded to an area of approximately 0.15 mm x 0.15 mm. None of these
cases contained metal backed acetabular cups or had an acetabular rim projection with an
eccentricity higher than 0.98. Additionally all images contained right facing necks, with
any images in the dataset containing left facing necks mirrored. They were cropped to a
600 x 600 pixels region of interest approximately centered on the femoral head. The FFD
and OFD were unknown for all of the radiographs, and the best, expected and worst case
scenarios stated by Krismer et al. [59] were evaluated. The effect of these values on was
investigated.
As the beam centre was unknown the centre of the acetabular rim ellipse was used as
the beam centre. The ellipse parameters were thus converted to millimetres and normalized
by the FFD to allow to be recovered. was constrained to lie between 0 and . These
normalized ellipse parameters were then discarded and the value of was used to compute
a new rim centre as per Equation 6.3, with ellipse parameter a used as r. Examples of the
corrected centres can be seen in Figures 6.7 and 6.8.
The distribution of rotations in depth is shown in Figure 6.9. The values do not change
if the FFD and OFD are varied. The absolute difference between the corresponding rotations in depth at year 1 and year 5 is shown in Figure 6.10. Additionally the radiographs
with the least rotation in depth and the most rotation in depth are shown in Figure 6.11.
92
(a)
(b)
(c)
Figure 6.7: Correction of centres where (a) FFD=1300 and OFD=180 (b) FFD=1000 and
OFD=225 (c) FFD=900 and OFD=270
93
(a)
(b)
(c)
Figure 6.8: Correction of centres where (a) FFD=1300 and OFD=180 (b) FFD=1000 and
OFD=225 (c) FFD=900 and OFD=270
94
(a)
(b)
Figure 6.9: Distribution of rotations in depth obtained for (a) Year 1 and (b) Year 5
radiographs
95
Figure 6.10: Distribution of the difference between rotations in depth obtained for Year 1
and Year 5
(a)
(b)
Figure 6.11: Radiograph with (a) lowest rotation in depth (53.2 ) and (b) highest rotation
in depth (85.9 ) from dataset
96
6.4
Summary and Discussion
This chapter has been concerned with correcting the eccentricity error, the distance between the centre of the projected ellipse and the centre of the circle that was projected
onto the radiograph. This chapter has described a method of correcting for this error by
estimating the rotation in depth of the circle necessary to generate the ellipse projected on
the radiograph. In the next chapter this method of correction shall be applied to actual
wear measurements and the effect of the correction on individual wear measurements shall
be examined.
In Table 6.1 it became apparent that the worst case values for eccentricity error could
be in excess of the total wear during the lifespan of a prosthesis. However the results show
that the difference between the values of rotation in depth at years 1 and 5 are between
0 to 15 . Therefore it is possible that most of the systematic eccentricity error will be
eliminated in duo-radiographic measurements. That is to say that the same approximate
amount of error between the actual and projected centres on the radiographs will be made
on both measurements, and thus eliminated in the calculation of wear. This is difficult
to prove however, given the uncertainty introduced into the pose estimation. In a worst
case the difference in year 1 and year 5 poses could actually be different by exactly one
quadrant (i.e. the first radiograph is rotated to 75 while the second is rotated to 105 ).
In the next chapter uncorrected acetabular wear is measured initially by the active
ellipses method, and then compared to corrected measurements to ascertain the overall
effect of the eccentricity error on the measurements.
97
Chapter 7
Measuring Acetabular Wear

This chapter describes using active ellipses to measure acetabular wear on sets of clinical
radiographs. It makes use of the initial estimation of the ellipse axes and orientation
parameters, the model of appearance and search method described in Chapter 4, and the
robust parameter estimation of Chapter 5 to localize the femoral head and acetabular rim
contours. Once both contours are localized the eccentricity error of the acetabular rim is
corrected for as described in Chapter 6.
However this is not enough to measure linear acetabular wear. As noted in Chapter 2
linear wear is usually quoted as a rate of wear over time. Thus it is necessary to consider
at least two points in time. In this chapter duo-radiographic measurements are made on
radiographs taken during the first and fifth years of the lifespan of a primary THR.
7.1
Experimental Parameters
In addition to the protocol laid down in Chapter 4 it is necessary to expand on how the
model of grey level was built that was used in these experiments. The training set that the
model was built from consisted of 45 postoperative radiographs containing Zimmer CPT
prostheses with a 22.225mm (or more intuitively
7
8
inches) diameter head. This model was
to be used on a test set consisting of 50 cases of Zimmer CPT prostheses with a 28mm
head diameter which had year 1 and year 5 standard clinical radiographs. This was the
same dataset described in Section 6.3 of Chapter 6.
On the training set the femoral head ellipse parameters were obtained by placing 9
98
equi-distant points on the contour of the femoral head, as shown in Figure 7.2(a). A
LSO fit was used to then generate the resultant ellipse parameters. Similarly, as shown
in Figure 7.2(b), 18 points were placed on the contour of the acetabular rim wire marker.
It should be noted that this is more than were placed on the femoral head contour as
the ellipse is more eccentric and thus more points were used to reduce curvature bias and
overestimation of eccentricity. In this case the points were not equi-distant, but rather
increased in frequency around the ends of the ellipse. The mean sub-axes were scaled by
the known diameter of the femoral heads in the test set. Thus if the method was trained on
a set of 22mm femoral heads but used to locate a 28mm femoral head the initial estimate
would be scaled by
28
22 .
The resultant mean femoral head and acetabular rim parameters
are visualised with initial centre coordinates provided by template matching in Figure 7.1.
7.2
Uncorrected Wear Experiment
A typical method of assessing the performance of an ASM is comparing the landmark

points on the localized contour to those on a contour that has been annotated by an
expert. Deviation from this expert annotation is taken as a measure of error. Repeatability
and limits of agreement were measured. Given the Monte-Carlo approach to generating
minimal subsamples for robust parameter estimation detailed in Chapter 5 the active
ellipses method should exhibit some variation.
Repeatability is the variation in measurements obtained when multiple measurements
are made using the same instrument and techniques on the same parts or items. It is desirable for a system for measuring acetabular wear to have a low variation in measurements.
Repeatability is not the same as computing accuracy, where the amount of wear present
in each case is known, and thus measurements by the new system are compared to the
known ground truth. Unfortunately no such ground truth was available, so repeatability
was calculated instead.
Wear measurements were obtained automatically using the active ellipses method and
by a labour-intensive method requiring a human annotator to interactively select points
on the desired contours.
99
(a)
(b)
Figure 7.1: Average parameters visualised on a radiograph for (a) the femoral head and
(b) the acetabular rim
100
Parameter
Description
Value
k1
Length of femoral head model profile
31
k2
Length of rim model profile
31
m1
Length of femoral head search normals
111
m2
Length of rim search normals
61
N1
No. of femoral head ellipse normals
400
N2
No. of acetabular rim ellipse normals
400
0.1 pixels
0.1 pixels
xc
0.1 pixels
yc
0.1 pixels
10
Maximum number of iterations before termination
30
LSO step size
1.2
300
EOF
Error of fit function used for LMedS fitting
EOF4
Table 7.1: Variables for the active ellipses used in obtaining wear measurements
7.2.1
Active Ellipses
The initial estimate of the femoral head ellipse was obtained as described in Chapters 4
and Chapter 5, and from this initial estimate the desired contours were localized. The
femoral head was localized first, and its centre was used as the initial centre parameters
of the acetabular rim, while a, b and were set to the mean rim parameters learned from
the training set and then ran until convergence. With the uncorrected centroids of the
femoral head and acetabular rim ellipses located in both the year 1 and year 5 radiographs
a measurement of wear could be obtained.
The values for the active ellipse variables that were used during this search are summarised in Table 7.1. Since the initial mean acetabular rim ellipse parameters were not
always very accurate due to variability in the projection of the rim wire, the method
101
checked for implausible parameters as an indication of fitting failure. In such cases, the
active ellipse reinitialised itself by reverting to the initial ellipse parameters and continuing the search. This process was repeated until either convergence or a fixed number of
iterations was completed, as shown in Table 7.1.
102
7.2.2
Comparative Method
The comparative method required the annotation of 27 points by the user (9 upon the
contour of the femoral head and 18 upon the contour of the acetabular rim as shown in
Figure 7.2). From these points the two sets of ellipse parameters were generated via LSO
ellipse fitting. The resultant ellipses were displayed over the radiograph and the user had
the choice of altering the points, accepting the fit, or restarting their annotation. This is
similar to the protocol used for creating the model of appearance described in Chapter 4.
Once the ellipse centroids had been obtained on both year 1 and year 5 radiographs wear
could be computed.
Four different annotators used this comparative labour-intensive method: an information technology professional, a podiatrist, an orthopaedic secretary and an orthopaedic
surgeon. Each annotator was familiarised with the tasks by performing measurements on
20 postoperative radiographs that were not included in the test set. They were asked to
annotate the test set of radiographs twice, at their own pace, but with a gap of at least
one day between repeat annotations.
(a)
(b)
Figure 7.2: (a) A set of data points on the femoral head. (b) A set of data points on the
acetabular rim.
103
7.2.3
Measuring Wear
(a)
(b)
(c)
Figure 7.3: (a) Vector obtained at Year 1 (b) Vector obtained at Year 5 (c) Wear computed
between the two vectors
On each radiograph a vector extends from the femoral head centre to the acetabular
rim centre on each image. The co-ordinate system of this displacement vector is such
that the femoral head centroid is at the origin as shown in Figure 7.3. Thus there is one
vector found on the year 1 radiograph and one on the year 5 radiograph. Conversion of
104
the vectors from pixels to millimeters is trivial given the known diameter of the femoral
head as a correction factor can be computed as follows
c=
d
ah + bh
(7.1)
where c is the correction factor that the value in pixels should be multiplied with to give
a value in mm, d is the known diameter of the femoral head in mm and ah and bh are the
major and minor semi-axis of the femoral head ellipse in pixels.
However it is worth noting that if the system generated an erroneous femoral head
ellipse the error would propagate further for any calculation using converting from pixels
to millimeters. Fortunately the femoral head is easily located in the image given a good
initialisation, and thus this problem does not occur.
7.2.4
Statistics
The methods of Bland and Altman [7] were used to measure repeatability and compare
the performance of the active ellipses method with the method described in Section 7.2.2.
They described a method for comparing one measurement technique with another to see
whether they agree sufficiently to be interchangeable. Standard deviations were obtained
for the difference of paired duo-radiographic measurements of wear. Additionally standard
deviations were obtained for paired uni-radiographic measurements of the year 1 and year
5 radiographs, taking the Euclidean distance between the femoral head centre and the
acetabular rim centre as a measurement in millimeters. These standard deviations measured the repeatability of each of the methods, where a high standard deviation indicates
an undesirable and high variation between two paired measurements. A mean that is not
close to 0 indicates a bias between the paired measurements.
Limits of agreement between the active ellipses and each of the annotators were calculated. The standard deviation of the difference between corresponding measurements
made by each of the methods was calculated. For experiments where there is more than
one measurement per case the mean of the measurements may be used instead. The limits of agreement were thus defined by the resultant mean and two standard deviations
allowing a 95% confidence interval and are shown in Table 7.2. Bland and Altman plots,
105
that is plots of the mean of both paired measurements or measurement means against the
difference, were obtained. These can be seen in Figures 7.4-7.7.
The actual wear measurements are summarized in table 7.6 and a distribution of
wear has been plotted for the mean measurements of the active ellipse method and the
annotators. These can be seen in Figures 7.8-7.12.
Figure 7.4: Bland and Altman plots for Annotator 1 and the Active Ellipse Method
106
107
Figure 7.8: Distribution of wear over 4 years for the active ellipses method
108
Figure 7.9: Distribution of wear over 4 years for annotator 1
109
Method
Upper
Lower
Limits
Limits
Annotator 1
0.8
-0.55
Annotator 2
0.3
-0.56
Annotator 3
0.35
-0.54
Annotator 4
1.09
-2.08
Mean
0.64
-0.93
Trimmed Mean
0.33
-0.55
Table 7.2: Limits of agreement of each annotator with the active ellipses method, mean
limits of agreement and alpha trimmed mean where = 0.5
110
Figure 7.12: Distribution of wear over 4 years for annotator 4. Outlying measurements
are not shown here
7.2.5
Discussion
Annotator 4s high standard deviation for measuring wear (1.42 mm) was due to erroneous annotations of radiographs. These can clearly be shown in Figures 7.7 and 7.13.
Method
Mean
Standard Deviation
(mm)
(mm)
Annotator 1
0.02
0.22
Annotator 2
0.02
0.31
Annotator 3
0.05
0.33
Annotator 4
-0.18
1.42
Annotator 4
0.05
0.42
0.01
0.11
(Trimmed)
Active Ellipse
Table 7.3: Mean and standard deviation of differences between paired wear measurements
with no eccentricity correction
111
Figure 7.13: Distribution of wear over 4 years for annotator 4 showing outlying measurements not shown in Figure 7.12
An example of these outliers is shown in Figure 7.14. As the standard deviation of the
distance between centroids was highest at year 5 (1.47mm) this is where the errors occurred. The annotator accepted the fit knowing it was incorrect. Even with an alpha
trim (=0.1) of the wear measurements Annotator 4 remains the most variable of the
Method
Mean
Standard Deviation
(mm)
(mm)
Annotator 1
0.04
0.24
Annotator 2
0.00
0.33
Annotator 3
0.07
0.20
Annotator 4
0.01
0.46
Active Ellipse
0.00
0.09
Table 7.4: Mean and standard deviation of centroid displacements in Year 1 radiographs
112
Method
Mean
Standard Deviation
(mm)
(mm)
Annotator 1
-0.01
0.19
Annotator 2
0.00
0.29
Annotator 3
-0.01
0.30
Annotator 4
-0.26
1.47
Annotator 4
-0.21
0.53
0.00
0.07
(Trimmed)
Active Ellipse
Table 7.5: Mean and standard deviation of centroid displacements in Year 5 radiographs
Method
Linear
Rate of
Wear
Linear
(mm)
Wear (mm/yr)
Active Ellipse
0.42 0.32
0.11 0.08
Annotator 1
0.30 0.16
0.08 0.04
Annotator 2
0.55 0.33
0.14 0.08
Annotator 3
0.52 0.33
0.13 0.08
Annotator 4
0.92 1.07
0.23 0.27
Table 7.6: Observed Parameters of 2-D Wear

annotators. The high standard deviation of wear measurements indicated the annotator
could not confidently measure acetabular wear over the year 1 to year 5 period.
It is interesting to note that the standard deviations of uniradiographic measurements,
as shown in Tables 7.4 and 7.5 using the Active Ellipses method are similar for both the
year 1 and year 5 radiographs, while those of the human annotations tend to have a large
difference between standard deviations of uniradiographic measurements, most notably
differences of 0.05 mm, 0.1 mm and 1.01 mm respectively for Annotators 1, 2 and 4. It
appears the majority of the annotators exhibited a greater variance on the year 5 dataset.
113
Figure 7.14: An incorrect measurement performed on a year 5 radiograph by annotator 4

The standard deviation of the active ellipses method is sufficient to detect the expected
wear over a period of 4 years (year 1 to 5) where wear manifests itself as a mean displacement of 0.4mm to 0.8mm. It could conceivably measure over a shorter period such as
year 1 to year 3. The standard deviations of the annotators indicate that it could not be
applied to measuring wear over a shorter period of time (such as year 1 to year 2) as the
standard deviation would be equal to the expected wear mean.
The limits of agreement of the active ellipses and Annotators 1, 2 and 3 are just within
clinically relevant ranges, namely between four years of wear at the expected rates given
in Chapter 2- 0.36mm and 0.92mm. The reason for the limits of agreement being so large
is the relative difference in repeatabilities between the human annotators and the active
ellipse method. When comparing a system with poor repeatability to a system with good
repeatability the limits of agreement tend to be large.
The mean annual wear of Annotators 1, 3 and 4, shown in Table 7.6 were close to
that of the active ellipses, while Annotator 2 overestimated wear at 0.23 mm/year due
incorrectly annotated measurements. Annotator 1 appears to underestimate the amount
of wear. Annotator 1s distribution of wear, shown in Figure 7.9 is quite different to that
of the active ellipses as shown in Figure 7.8. This explains the high limits of agreement
between Annotator 1 and the active ellipse method shown in Table 7.6 despite the fact the
mean wear is close. On the other hand Annotator 2 and 3 have similar wear distributions,
114
shown in Figures 7.10 and 7.11, to that of the active ellipse method shown in Figure 7.8.
This is also reflected in the closer limits of agreement in Table 7.6.
7.3
Normals
In this experiment the number of normals extended from the active ellipses was varied.
The effect of varying this parameter on the repeatability of the active ellipses method
was examined. Paired acetabular wear measurements were obtained on the test data of 50
cases using the protocol described in Section 7.2 and the variables shown in Table 7.1 were
used with the exception of N1 and N2 , the number of normals for each of the ellipses. The
methods of Bland and Altman were used, this time to estimate the mean and standard
deviations for sets of wear measurements using differing numbers of normals. These are
summarised in Table 7.7 and Figure 7.15.
Number of Normals
Mean
Standard Deviation
(N1 and N2 )
(mm)
(mm)
40
0.02
0.16
80
0.01
0.14
120
-0.01
0.12
160
-0.01
0.12
200
0.01
0.11
400
0.00
0.11
800
0.00
0.11
Table 7.7: Mean and standard deviation of uni-radiographic measurements performed on

differing resolutions.
From these results it becomes apparent that the benefit of increasing the number of
normals extended from the ellipses lessens as the number of normals is increased. It was
decided that the best number of normals was 400, where all bias has been removed (i.e.
the mean of the difference between paired measurements is within two 2 decimal places of
0) and the standard deviation is lowest. Doubling the number of normals to 800 appears
115
Figure 7.15: Plot of normals extended from an active ellipse against repeatability of paired
measurements of acetabular wear
to have no additional benefit.
7.4
Resolution
In this experiment the resolution at which the radiograph was scanned was varied. Two
datasets were created by scanning 20 radiographs containing 28 mm Zimmer CPT prostheses. One set of radiographs was scanned at 150 dpi and the other scanned at 300
dpi. Uncorrected uni-radiographic measurements were measured using the active ellipses
method and by one annotator using the more labour intensive method discussed in Section 7.2.2. The active ellipse method required a model of appearance to be built for each
resolution. Due to the small size of the dataset leave-1-out was used for the searches conducted. The model was built using the 19 remaining images and the 150 dpi search used
the parameters described in Table 7.1 while the 300 dpi used those described in Table 7.9.
Increasing the resolution of the scans to 300 dpi lowers the standard deviation for
both methods. This is due to the fact that pixels correspond to a smaller area. On
150 dpi a pixel corresponds to 0.15mm approximately while at 300 dpi it corresponds to
0.07mm approximately. Thus decreasing the area corresponding to a single pixel lowers
the variation between the paired measurements.
116
Method and Resolution
Mean
Standard Deviation
(mm)
(mm)
Annotator (150 dpi)
-0.12
0.21
Annotator (300 dpi)
-0.08
0.09
Active Ellipses (150 dpi)
0.02
0.09
Active Ellipses (300 dpi)
0.02
0.05
Table 7.8: Mean and standard deviation of uni-radiographic measurements performed on

differing resolutions
7.5
Corrected Wear Experiment
As noted in Chapter 6 the centre of the acetabular rim is not the centre of the ellipse that
appears on the radiograph. In cases where the rotation in depth of the year 1 and year
Parameter
Description
Value
k1
Length of femoral head model profile
61
k2
Length of rim model profile
61
m1
Length of femoral head search normals
221
m2
Length of rim search normals
121
N1
No. of femoral head ellipse normals
400
N2
No. of acetabular rim ellipse normals
400
0.1 pixels
0.1 pixels
xc
0.1 pixels
yc
0.1 pixels
10
Maximum number of iterations before termination
30
LSO step size
1.2
300
EOF
Error of fit function used for LMedS fitting
EOF4
Table 7.9: Variables for the active ellipses used in the resolution experiment at 300 dpi
117
5 rims are not identical the eccentricity error may prove to be significant. Therefore it
is necessary to examine the amount of correction needed, and the effect this has on the
wear measurements. Given that the Active Ellipses method was shown to be the most
repeatable of the methods examined in the previous experiment the ellipses generated
are used to demonstrate the effect of correcting for eccentricity upon a series of wear
measurements.
7.5.1
Correction Parameters
A further three sets of wear calculations were obtained using eccentricity error correction
of the acetabular rim centre. The first case considered was when the FFD was set to the
maximal value stated in the literature (1300 mm) while the OFD (180 mm) was set to the
minimal value. Thus the eccentricity error occurring with these parameters was smallest.
In the second case however the FFD is minimal (900 mm), while the OFD (270 mm) is
maximal. The eccentricity error obtained using these parameters was largest. The third
case used the most commonly stated FFD (1000mm). However no such value is commonly
stated for the OFD so the value used is the mean of the upper and lower bounds (225mm).
7.5.2
Statistics
In addition to recomputing the wear parameters and using the methods of Bland and Altman to assess the repeatability of paired and corrected wear measurements, the resultant
wear value was subtracted from the original wear value and the distribution was plotted.
This distribution is shown in Figure 7.16. A mean and standard deviation for this error
was additionally estimated and can be seen in Table 7.12.
7.5.3
Discussion
Tables 7.10 and 7.11 show that eccentricity correction has no overall effect on the repeatability of the duo-radiographic method or on the mean of the wear measurements obtained.
However examining the difference between uncorrected and corrected wear measurements,
as shown in Table 7.12 and Figure 7.16 reveal this is because the mean of the error is close
to zero for all of the parameter sets. However the eccentricity error can have a noticeable
118
Figure 7.16: Distribution of differences between uncorrected and corrected wear values
Method
Mean
Standard Deviation
(mm)
(mm)
No Correction
0.02
0.11
Upper Bounds
0.01
0.10
Mean Bounds
0.02
0.10
Lower Bounds
0.02
0.11
Table 7.10: Mean and standard deviation of differences between paired active ellipse wear
measurements with varying eccentricity correction
Method
Linear
Rate of
Wear
Linear
(mm)
Wear (mm/yr)
No Correction
0.42 0.32
0.11 0.08
Upper Bounds
0.43 0.33
0.11 0.08
Mean Bounds
0.43 0.33
0.11 0.08
Lower Bounds
0.44 0.33
0.11 0.08
Table 7.11: Observed parameters of 2D Wear with eccentricity correction
119
Method
Mean
Standard Deviation
(mm)
(mm)
Upper Bounds
0.00
0.02
Mean Bounds
-0.00
0.03
Lower Bounds
-0.02
0.04
Table 7.12: Mean and standard deviation of difference between uncorrected and corrected
wear measurements
effect on individual measurements, the maximums in the upper, mean and lower bounds
being an error of 0.04mm, 0.08mm and 0.10mm respectively. The standard deviation of
this error, quoted in Table 7.12 is as high as the standard deviation of the active ellipses
method. Clearly these values are clinically relevant as they amount to 1 year of wear given
the assumed wear rates described in Chapter 2. They also highlight the importance of
recording the FFD at exposure, given an uncertainty of 0.06mm for the maximum case.
This is the first method that the author is aware of that uses eccentricity correction
of ellipses to locate the centre of the acetabular rim. Methods such as EBRA [59] or
that of Eggli et al. [31] that use the centre of the acetabular rim ellipse that appears on
radiographs as a reference point do not make any attempt to correct for the eccentricity of
the acetabular rim. One of the reasons why the eccentricity error is relatively low across
the dataset in this chapter is because the rotation in depth is similar for the year 1 and year
5 radiograph so the centre point of the acetabular rim should remain consistently biased.
On the other hand the other methods discussed are uni-radiographic measurements and
thus the eccentricity error is likely to be higher for their measurements.
7.6
Summary
In this chapter wear measurements made by four individuals using a labour-intensive

method involving annotation of digitized radiographs were compared to wear measurements made by the active ellipses method. The methods were tested on images of fifty
clinical cases with radiographs taken at years 1 and 5. The automated method was found
to have lowest variation between repeated measurements and limits of agreement between
120
the methods indicated that the active ellipses replace the labour intensive method. Further experiments showed that increasing the resolution to 300 dpi lowers the standard
deviations of both active ellipses and the labour intensive method, while the number of
normals extended from the ellipses that were used in these experiments was optimal.
Additionally eccentricity error correction of the active ellipses method showed the effect
of the eccentricity error was significant on wear measurements, altering the uncorrected
wear values by up to 0.1mm depending on the exposure parameters.
121
Chapter 8
Conclusions
The main objective of this thesis, that is to create a new system for assessing hip implants
that should be as automated as possible without requiring specialist equipment has been
met. The method was more repeatable than a human annotator at the same task. The
standard deviation of the difference between paired measurements was 0.11mm while
the closest human annotator was 0.22mm and the furthest was 1.42mm. The limits of
agreement between the human annotators were found to be just within clinically relevant
ranges. This is due to the fact that a method with high repeatability often has poor
agreement with one with low repeatability. In this instance increasing the automation
of the measurements in radiographs substantially increases the amount of data points
available to fit ellipses to. The number of data points found on each of the contours by
the new method cannot be obtained by human observers without an extended period of
labour. However the accuracy of the system was not examined, and this is explained below
with suggestions for future work.
This thesis has included topics such as modelling, robust parameter estimation, eccentricity error correction and measuring wear. The investigation into automated assessment
of AP radiographs has led to several other novel developments worth noting.
The robust fitting algorithm proposed is the first to provide data comparing the orthogonal distance directly to previous EOF functions. It proves to be a good choice of
EOF function, with the foci bisector distance being computationally cheaper with a similar performance. It is also the first time the LMedS algorithm has been applied in an
122
application such as this.

The magnitude of the systematic eccentricity error introduced by using the projected
acetabular rim centre as a reference point is sufficient to warrant further investigation.
Using the lower bound parameters for FFD and OFD this error was found to be at most
0.1mm. This error is potentially higher when one considers the ambiguity of the pose
estimation, where the rotation in depth could also equal . Correcting for this source
of error in the proposed method is novel and should also be applied to other methods that
use the centre of the rim. Examples of these methods include EBRA [59] and Eggli et
al. [31] when using the centre of the acetabular rim as a reference point.
In the development of the method described in this thesis several ideas for future
improvements and expansion have presented themselves. For instance in this thesis the
repeatability of paired measurements is measured in Chapter 7, not the accuracy of individual wear measurements. To measure this a ground truth measurement of wear is
required for each case investigated. Such ground truth was not available for these experiments but for future work a wear simulator could be used to produce radiographs
where the amount of acetabular wear in the AP plane is known. These could then be
compared with measurements obtained using the method to obtain a standard deviation
that measures reproducibility and thus validates the method. It would also be desirable to
expand the method to measure wear in the ML plane and thus combine AP and ML wear
measurements to measure 3-D wear. It would also be interesting to perform a detailed
comparison of the novel method and the other methods reviewed in Chapter 3.
The system finds a vector of wear with respect to the image plane but makes no
attempt to find the vector relative to the prosthesis. Detection of the stem axis has
already been successfully researched by Wigderowitz et al.[98], though their system was
originally designed to measure the longitudinal axis of the wrist bone. This system itself
could be extended by applying the model of appearance and using a more robust method
such as LMedS to fit the lines necessary to localise the stem axis. With the axis successfully
located the wear vector could be computed relative to the stem axis.
The method does not work on non-elliptical projections of the acetabular rim, for
example where the acetabular rim ellipse is so eccentric it is projected as a line. The
method has been used exclusively on non-metal backed cemented acetabular components.
123
A logical extension of this would be to apply the method to cementless and metal backed
components, though occlusion of the femoral head by the metal shell of the acetabular
component could make location of the head difficult. Additionally the performance of the
system could be investigated using the edge wire marker of the acetabular cup in cemented
prostheses instead of the rim wire marker. This would allow wear to be measured using
the minimum arc separation, or thinness between the femoral head contour and edge
wire marker. The same model of appearance could be applied to the edge wire marker.
Knudsen et al. [57] showed that using the thickness can improve the repeatability of a
method.
The model of appearance could be applied to measuring radiolucency. By building a
model of appearance around the stem for normal and radiolucent cases it may be possible
to detect radiolucency around the contours of the prosthesis. If the radiolucent model
better matches a profile around the prosthesis than the normal model that profile could
be considered radiolucent.
To fully assess the effect of the eccentricity error without the simplification made in
Chapter 6 a model could be produced whereby all the parameters are known and thus
the eccentricity error could be obtained by using the original Equations 6.1 and 6.2. A
model of likely positions of the beam centre on the symphysis pubis could also be created,
thus taking into account the translation of the circular object when computing eccentricity
error. It would also be desirable to accurately segment the symphysis pubis by means of
an ASM and use the centre of the model as the beam centre. However this in itself is a
project of significant scope.
The method for initialising the local search could certainly be improved on. Despite
its relative simplicity the normalized cross-correlation performed well, but a more representative template based on several training examples may reduce the number of failed
initialisations presented in Chapter 5. One method that presents itself would be to create
a mean and covariance matrix of the region of interest from the training set. When a
novel radiograph requires initialisation the Mahalanobis distance could then be computed
for each possible translation, rather than computing the normalized cross-correlation.
In conclusion this thesis describes a completely automated method for measuring wear.
The only other method reviewed in Chapter 3 able to make this claim is RSA-CMS. Unlike
124
RSA-CMS this method requires only a standard PC setup, and the means to digitize AP
radiographs. There is no need for specialised equipment or additional invasive apparatus.
This low overhead, coupled with the increased repeatability over manual annotation, makes
its use in the standard clinical setting more feasible.
125
Appendix A
The Ellipse
An ellipse is a conic section formed when a plane intersects a cone, but the plane is not
parallel with cones axis. This is shown in Figure A.1. It can also be defined as the locus
of points on a plane that have the same sum of distances from two given foci. If these foci
are coincident then the ellipse is circular.
Conversion between the different parametric forms is convenient in order to fully exploit
the characteristics of the ellipse. For example in order to calculate the algebraic distance
one can use the implicit form, while to calculate the foci bisector distance one must use
the explicit form. In order to estimate the 3D pose of a circle the symmetric matrix of
the projected ellipses coefficients can be used. Therefore in this appendix a summary of
these three representations and the methods of converting between them are described.
Figure A.1: Projecting both a circle and an ellipse from a circle
126
A.1
Explicit Form
The explicit form of an ellipse is the easiest to visualise. It consists of 5 parameters as

shown in Figure A.2. The major sub-axis, a, is the longest line from the centre of the
ellipse to the contour. The minor sub-axis, b, is orthogonal to the major sub-axis and is
the shortest line from the centre of the ellipse to the contour.
Figure A.2: Visualisation of the 5 ellipse parameters
Figure A.3: Canonical example of an ellipse, where xc = yc = = 0.

The canonical example of an ellipse is one that is centred at the origin with its major
sub-axis, is in the x-axis direction and its minor sub-axis is in the y-axis direction. This
is shown in Figure A.3. Any points on this ellipse satisfy the condition
127
x2 y 2
+ 2 = 1.
a2
b
(A.1)
If the ellipse is translated so the centre lies at (xc , yc ) the points on the ellipse satisfy
(x xc )2 (y yc )2
+
= 1.
a2
b2
(A.2)
Lastly, the ellipse may be rotated anti-clockwise by an angle . If the ellipse is both
translated and rotated then the condition all points along the ellipses satisfies becomes
((y yc ) sin () + (x xc ) cos ())2 ((y yc ) cos () (x xc ) sin ())2

+
=1
a2
b2
(A.3)
These five parameters are the explicit form of an ellipse. The eccentricity of an ellipse, e,
is
s
e=
A.2
b2
a2
(A.4)
Implicit Form
It is also possible to represent an ellipse using the general equation of a conic section. The
general equation of a conic is
Ax2 + Bxy + Cy 2 + Dxz + Eyz + F z 2 = 0
(A.5)
An ellipse lies on a plane and therefore can be assumed to have a constant z value. It is
therefore simpler to normalize the co-ordinate system by z and thus simplify the above
equation to:
Ax2 + Bxy + Cy 2 + Dx + Ey + F = 0
(A.6)
where A, B, C, D, E and F are coefficients of the conic section. The conic section is an
ellipse if the following constraint is met
B 2 4AC < 0
128
(A.7)
whereas
B 2 4AC = 0
(A.8)
B 2 4AC > 0
(A.9)
represent a parabola and a hyperbola respectively. For cases where the described ellipse
is not rotated B = 0. In cases where there is no translation along the x-axis D = 0. For
no translation along the y-axis E = 0. An ellipse is circular when A = C.
A.3
Matrix Form
The matrix form is simply another way of expressing Equation A.5:
Ax2 + Bxy + Cy 2 + Dx + Ey + F = XT EX = 0
(A.10)
where X = [x, y, 1] and E is the real, symmetric matrix of the ellipse as follows:
E = B/2
B/2 D/2
C
D/2 E/2
A.4
E/2
(A.11)
Conversion
Following Otepka and Fraser[71], conversion between the explicit form and the implicit
form can be described as follows:
cos2 sin2
A =
+
b2
a2
= 2 cos sin 1 1
B
b2 a2
2
sin cos2
C =
+
b2
a2
yc2 C
F = 1 x2c A xc yc B
= 2xc A yc B
D
129
(A.12)
(A.13)
(A.14)
(A.15)
(A.16)
= 2yc C yc B
E
A
A=
F
B
B=
F
C
C=
F
D
D=
F
E
E=
F
(A.17)
F =1
(A.23)
(A.18)
(A.19)
(A.20)
(A.21)
(A.22)
Conversion from the implicit to the explicit form is as follows:
BE 2(CD)
4AC B 2
BD 2(AE)
yc =
4AC B 2
1
B
= tan1
2
AC
xc =
a=
s
b=
(A.24)
(A.25)
(A.26)
E
1 D
2 xc 2 yc
C cos2 B sin cos + A sin2
(A.27)
E
1 D
2 xc + 2 yc
A cos2 + B sin cos + C sin2
(A.28)
130
Appendix B
System Overview
This appendix gives a very high level overview of the method detailed in this thesis. It
breaks the system down into three subsystems: Training subsystem - where the user marks up a set of training images and the system
generates ellipse parameters for each image. These ellipse parameters are then used
to calculate the first order normalized grey level derivative. The mean and covariance
of all obtained profiles is then computed for use as a model of appearance.
Contour Localization subsystem - which locates the desired contours and optionally
performs eccentricity error correction.
Wear Calculation subsystem - which calculates wear given the femoral head and
acetabular rim parameters on two radiographs of the same THR taken at different
points in time.
Pseudocode and dataflow diagrams (Figures B.1-B.3) are provided for each subsystem
to aid in understanding the workings of the system. The reader is directed to chapters
within the thesis in order to gain an more thorough understanding of how a particular
module works.
B.1
Overall System
1. User uses training subsystem to build model of appearance and learn mean ellipse
parameters.
131
2. Ellipse contours are localized using active ellipses subsystem and the models of appearance. The acetabular rim centre optionally undergoes eccentricity error correction.
3. Repeat Step 2 until system has localized contours of paired radiographs taken at
different periods of time during the prosthesiss lifespan.
3. Use these localized contours from the same THR case to calculate wear.
B.2
Training Subsystem
Figure B.1: Data flow diagram of the training subsystem
132
B.2.1
Annotation Pseudocode
The annotation should be completed sequentially, with the femoral head completed first
so that the parameters can be used to reject acetabular rim profiles that pass through the
femoral head. Such acetabular rim profiles are not useful as the rim is occluded by the
femoral head and thus would perturb the mean and covariance.
Distinct pseudocode for femoral head and acetabular rim localization is provided below.
B.2.2
Femoral Head Annotation Pseudocode
Femoral Head Annotation

1. User select training image to load.
2. System displays selected training image.
3. User annotates points on desired contour.
Femoral Head Ellipse Fitting
4. System fits ellipse through datapoints using LSO Algorithm (see Chapter 4). Ellipse
parameters returned for later use (see below).
Femoral Head Profile Calculation
5. System extends pre-determined number N1 of normals from ellipse.
5.1 System ejects 25% of profiles from the bottom left quadrant of the ellipse.
6. System calculates normalized first order derivative and retain for later use.
RETURNS: A normalized first order derivative profile. Many normalized first order derivative profiles can be used to build a model as described below.
B.2.3
Acetabular Rim Annotation Pseudocode
When building the model for the acetabular rim the corresponding femoral head ellipse
parameters are required in order to discount any profiles obtained where the femoral head
occludes the acetabular rim.
133
Acetabular Rim Annotation

1. User selects training image to load.
2. System displays selected training image.
3. User annotates points on desired contour.
Acetabular Rim Ellipse Fitting
4. System fits ellipse through datapoints using LSO Algorithm (see Chapter 4).
Acetabular Rim Profile Calculation
5. System extends pre-determined number N2 of normals from ellipse.
5.1 System rejects profiles that pass through the femoral head ellipse.
6. System calculates normalized first order derivative of each profile, which is retained for
later use.
RETURNS: A normalized first order derivative profile. Many normalized first order derivative profiles can be used to build a model as described below.
B.2.4
Model Calculation Code
1. System calculates mean of all given profiles.

2. System calculates covariance of all given profiles.
3. System calculates mean axes and orientation.
RETURNS: Mean and covariance calculated from all given profiles and mean axes (
a, b)
and orientation ().
134
B.3
Contour Localization Subsystem
Template Matching
1. System obtains initial ellipse centre estimate (xc , yc ) using template matching (see
Chapter 5).
Note: In the event of the template match failing it is additionally possible for the user to
supply the starting centre estimates via a single mouse click.
2. System uses mean axes and orientation with centre estimates to generate initial estimate
of all 5 ellipse parameters.
Femoral Head Contour Localization
3. System extends N1 normals from the ellipse to find the best matches for the model of
appearance. See Chapter 4 for more information.
4. System updates the parameters (a, b, xc , yc , ) to fit the located points via an LMedS
fit (see Chapter 5).
6. System uses average length of both major and minor axis in pixels and known femoral
head diameter to calculate pixel to millimetre ratio.
Acetabular Rim Contour Localization
7. System uses mean axes and orientation with centre estimates to generate initial estimate
of all 5 ellipse parameters.
8. System extends N2 normals to the ellipse to find the best matches for the model of
appearance.
9. System updates the parameters (a, b, xc , yc , ) to fit the located points via an LMedS
fit (see Chapter 5). Reject any points lying close to the femoral head contour.
135
Figure B.2: Data flow diagram of the Contour Localization subsystem and subsequent
eccentricity error correction
Eccentricity Error Correction
11. (optional) System performs eccentricity error correction on the acetabular rim centre.
(See Chapter 6 for more details)
136
RETURNS: Femoral head and acetabular rim ellipse parameters, pixel to millimetre conversion ratio, corrected acetabular rim centre (optional).
137
B.4
Wear Measurement Subsystem
Figure B.3: Data flow diagram of the wear calculation subsystem and subsequent eccentricity error correction
1. System calculate vector with origin coincident with the femoral head centre and extending to the acetabular rim centre for both sets of parameters. The acetabular centre can
either be the ellipse centre or calculated using the optional eccentricity error correction
step noted above.
2. System converts the vector values from pixels to millimetres.
3. System Calculate the distance between the two vectors.
RETURNS: Distance between the two input vectors as a measure of wear in millimetres.
If the number of years between the paired radiographs is known a rate of wear can be
calculated.
138
Appendix C
Publications resulting from this

work
The following papers have resulted from work done in this thesis.
S. Kerrigan, S.J. McKenna, I.W. Ricketts, and C.A. Wigderowitz. Measurement of acetabular wear using intelligent ellipses. British Orthopaedics Research Society (Cardiff),
page 35, 2002.
S. Kerrigan, S.J. McKenna, I.W. Ricketts, and C.A. Wigderowitz. Analysis of total hip
replacements using active ellipses. Medical Image Understanding and Analysis (Sheffield),
pages 177-180, 2003.
S. Kerrigan, S.J. McKenna, I.W. Ricketts, and C.A. Wigderowitz. The Effect of Automation on Repeatability of Measuring Acetabular Wear. British Orthopaedics Research
Society (Manchester), page 32, 2004.
S. Kerrigan, S.J. McKenna, I.W. Ricketts, and C.A. Wigderowitz. Eccentric Elliptical
Contours in Total Hip Replacements. Medical Image Understanding and Analysis (London), pages 216219, 2004.
S. Kerrigan, S.J. McKenna, I.W. Ricketts, and C.A. Wigderowitz. Eccentricity Error
Correction for Automated Estimation of Polyethylene Wear after Total Hip Arthroplasty.
139
Submitted to Medical Image Understanding and Analysis (Bristol), 2005.

S. Kerrigan, S.J. McKenna, I.W. Ricketts, and C.A. Wigderowitz. Radiographic Measurement of Polyethylene Wear after Total Hip Arthroplasty: A Comparison of Two Methods
with Varying Levels of Automation. Submitted to the Journal of Orthopaedic Research,
2005.
140
Bibliography
[1] S. J. Ahn, W. Rauh, H. S. Cho, and H. Warnecke. Orthogonal Distance Fitting of
Implicit Curves and Surfaces. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 24(5):620638, 2002.
[2] S. J. Ahn, W. Rauh, and H. J. Warnecke. Least Squares Orthogonal Distances
Fitting of Circle, Sphere, Ellipse, Hyperbola and Parabola. Pattern Recognition,
34:22832303, 2001.
[3] S.J. Ahn and H.J. Warnecke. Systematic Geometric Image Measurement Errors of
Circular Object Targets: Mathematical Formulation and Correction. Photogrammetric Record, 16:485502, 1999.
[4] J. Ball and T. Price. Chesneys Radiographic Imaging Sixth Edition. Blackwell
Science, 1995.
[5] R. Biedermann, M. Krismer, B. Stockl, P. Mayrhofer, E. Ornstein, and H. Franzen.
Accuracy of EBRA-FCA in the measurement of migration of femoral components
of total hip replacement. Einzel-Bild-Rontgen-Analyse-femoral component analysis.
Journal of Bone and Joint Surgery, 81-B:266272, 1999.
[6] A. Blake and M. Isard. Active Contours. Springer, 1998.
[7] J.M. Bland and D.G. Altman. Statistical Methods for Assessing Agreement Between
Two Methods of Clinical Measurement. Lancet, i:307310, 1986.
[8] A.W. Blom, A.H. Taylor, G. Pattison, S. Whitehouse, and G.C. Bannister. Infection
after Total Hip Arthroplasty: The Avon Experience. Journal of Bone and Joint
Surgery, 85-B:956959, 2003.
141
[9] R.C. Bolles and M.A. Fischler. A RANSAC-Based Approach to Model Fitting and
its Applications to Finding Cylinders in Range Data. International Joint Conference
on Artificial Intelligence, 7:637643, 1981.
[10] F.L. Bookstein. Fitting conic sections to scattered data. Computer Graphics and
Image Processing, pages 5671, 1979.
[11] M. Bould, S. Barnard, I.D. Learmonth, J.L. Cunningham, and J.R.W Hardy. Digital
Image Analysis: Improving Accuracy and Reproducibility of Radiographic Measurement. Journal of Clinical Biomechanics, 14:434437, 1999.
[12] K. Burckhardt, C.H. Gerber, J. Hodler, Notzli H., and G. Szekely. Precision of
Distance Determination Using 3D to 2D Projections: The Error of Migration Measurement Using X-Ray Images. Medical Image Analysis, 4:375388, 2000.
[13] J.J. Callaghan, D.R. Pedersen, J.P. Olejniczak, D.D. Goetz, and R.C. Johnston.
Radiographic Measurement of Wear in 5 Cohorts of Patients Observed for 5 to 22
Years. Clinical Orthopaedics and Related Research, 317:1418, 1995.
[14] H.E. Cates, P.M Faris, E.M Keating, and M.A. Ritter. Polyethylene Wear in Cemented Metal-Backed Acetabular Cups. Journal of Bone and Joint Surgery, 75B:249253, 1993.
[15] J. Charnley and Z. Cubic. Nine and Ten Year Results of Low-Friction Arthroplasty
of the Hip. Clinical Orthopaedics and Related Research, 95:925, 1973.
[16] J. Charnley and D.K. Halley. Rate of Wear in Total Hip Replacement. Clinical
Orthopaedics and Related Research, 112:170179, 1975.
[17] K. Chen and J. Wu. Measurement of Polyethylene Wear - A New Three-Dimensional
Methodology. Computer Methods and Programs in Biomedecine, 68:117127, 2002.
[18] J.C. Clarke, K. Black, C. Rennie, and H.C. Amsutz. Can Wear in Total Hip Arthroplasties Be Assessed from Radiographs? Clinical Orthopaedics and Related Research,
pages 126142, 1976.
142
[19] J.P. Collier, M.B. Mayor, V.A. Surprenant, L.A. Dauphinais, and R.E. Jensen. The
Biomechanical Problems of Polyethylene as a Bearing Surface. Clinical Orthopaedics
and Related Research, 261:107113, 1990.
[20] T. Cootes and C. Taylor. Statistical Models of Appearance for Medical Image Analysis and Computer Vision. Proceedings of the SPIE Medical Imaging, 4322:236248,
2001.
[21] T.F. Cootes, A. Hill, C.J. Taylor, and J. Haslam. The Use of Active Shape Models
for Locating Structures in Medical Images. Image and Vision Computing, pages
355366, 1994.
[22] T.F. Cootes and C.J. Taylor. Statistical Models of Appearance in Computer Vision.
University of Manchester, 2001.
[23] Zimmer Corporation. CPT 12/14 Hip System. Zimmer Medical Zone, 2004.
[24] Zimmer Corporation. ZCA Cup. Zimmer Medical Zone, 2004.
[25] H.R. Cowell. Editorial: Radiographic Measurements and Clinical Decisions. Journal
of Bone and Joint Surgery, 72-A:319, 1990.
[26] D.L. de Ipina. Video-Based Sensing for Wide Deployment of Sentient Spaces. Proceedings of the 2nd International Workshop on Ubiquitous Computing and Communications, 2, 2001.
[27] J.C. DeLee and J. Charnley. Radiological Demarcation of Cemented Sockets in Total
Hip Replacement. Clinical Orthopaedics and Related Research, 121:2032, 1976.
[28] P.A. Devane, R.B. Bourne, C.H. Rorabeck, R.M. Hardie, and J.G. Horne. Measurement of Polyethylene Wear in Metal-Backed Acetabular Cups. Clinical Orthopaedics
and Related Research, 319:303316, 1995.
[29] Draftware Developers. Polyware Pro 3D. 1995.
[30] M. Dickob, J. Bleher, and W. Puhl. Standardized analysis of acetabular cup migration in hip endoprosthesis using digital image processing. Unfallchirurg, 97:9297,
1994.
143
[31] S. Eggli, S. ZBrun, C. Gerber, and R. Ganz. Comparison of Polyethylene Wear with
Femoral Heads of 22mm and 32mm. A Prospective Randomised Study. Journal of
Bone and Joint Surgery, 84-B:447451, 2002.
[32] M.A. Fischler and R.C. Bolles. Random Sample Consensus: A Paradigm for Model
Fitting with Applications to Image Analysis and Automated Cartography. Communications of the Association for Computing Machinery, 24(6):381395, 1981.
[33] A. Fitzgibbon and R. Fisher. A Buyers Guide to Conic Fitting. Proceedings of the
British Machine Vision Conference, pages 513522, 1995.
[34] A.W. Fitzgibbon, M. Pilu, and R.B. Fisher. Direct Least Square Fitting of Ellipses.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 21:476480, 1999.
[35] D. Forsyth, J.L. Mundy, A. Zisserman, C. Coelho, A. Heller, and C. Rothwell.
Invariant Descriptors for 3-D Object Recognition and Pose. IEEE Pattern Analysis
and Machine Intelligence, 13(10):971991, October 1991.
[36] A. Furnes, S.A. Lie, L.I. Havelin, L.B. Engesaeter, and S.E Vollset. The Economic
Impact of Failures in Total Hip Replacement Surgery: 28,997 cases from the Norwegian Arthroplasty Register, 1987-1993. Acta Orthopaedics Scandinavia, 67(2):115
121, 1996.
[37] W. Gander, G.H. Golub, and R. Strebel. Least-Squares Fitting of Circles and Ellipses. Numerical Analysis, pages 6384, 1996.
[38] R.G.T. Geesink. Hydroxyapatite-Coated Total Hip Prostheses: Two-Year Clinical and Roentgenographic Results of 100 Cases. Clinical Orthopaedics and Related
Research, 261:3958, 1990.
[39] A.S. Greenwald and J.P. Garino. Alternative Bearing Surfaces: The Good, The Bad
and the Ugly. Journal of Bone and Joint Surgery, 83-A:6872, 2001.
[40] M.J. Griffith, M.K. Seidenstein, D. Williams, and J. Charnley. Socket Wear in
Charnley Low Friction Arthroplasty of the Hip. Clinical Orthopaedics and Related
Research, 137:3747, 1978.
144
[41] T.A. Gruen, G.M McNeice, and H.C. Amstutz. Modes of Failure of Cemented Stemtype Femoral Components. Clinical Orthopaedics and Related Research, 141:1726,
1979.
[42] R. Halir and J. Flusser. Numerically Stable Direct Least Squares Fitting Of Ellipses,
1998.
[43] K. Hardinge, M.L. Porter, P.R. Jones, D.W.L. Hukins, and C.J. Taylor. Measurement of Hip Prostheses Using Image Analysis: The Maxima Hip Technique. Journal
of Bone and Joint Surgery, 73-B:724728, 1991.
[44] W.H. Harris. Hybrid Total Hip Replacement: Rationale and Intermediate Clinical
Results. Clinical Orthopaedics and Related Research, 333:155164, 1996.
[45] F.N. Hatfield, R.M. Hall, R.J. King, E. Berry, P. Siney, and B.M. Wroblewski.
Image Analysis of Wear In Total Hip Replacements. Medical Image Understanding
and Analysis, 2001.
[46] W.A. Hoff, R.D. Komistek, D.A. Dennis, S. Walker, E. Northcut, and K. Spargo.
Pose Estimation of Artificial Knee Implants in Fluoroscopy Images Using A Template
Matching Technique. Proceedings of 3rd Workshop on Applications of Computer
Vision, IEEE, 1996.
[47] D.W. Howie, B. Vernon-Roberts, R. Oakeshott, and B. Manthey. A Rat Model of
Resorption of Bone at the Cement-Bone Interface in the Presence of Polyethylene
Wear Particles. Journal of Bone and Joint Surgery, 70-A:257263, 1988.
[48] T.S. Huang, A.M. Bruckstein, R.J. Holt, and A.N. Netravali. Uniqueness of 3D Pose
Under Weak Perspective: A Geometric Proof. IEEE Pattern Analysis and Machine
Intelligence, 17:12201221, 1995.
[49] J. Huber. Robust Statistics. Wiley, 1981.
[50] A. J. Hui, R. W. McCalden, J. M. Martell, S. J. MacDonald, R. B. Bourne, and
C. H. Rorabeck. Validation of Two and Three-Dimensional Radiograph Techniques
for Measuring Polyethelyne Wear After Total Hip Replacement. Journal of Bone
and Joint Surgery, 85-A:505511, 2003.
145
[51] M.H. Huo and S.M. Cook. Speciality Update: Whats New In Hip Arthroplasty.
Journal of Bone and Joint Surgery, 83-A:15981610, 2001.
[52] T. Ilchmann, M. Ljubisa, A. Joshi, K. Hardinge, J. Murphy, and H. Wingstrand.
Migration and Wear of Long-Term Successful Charnley Total Hip Replacements.
Journal of Bone and Joint Surgery, 80-B:377381, 1998.
[53] P.R. Jones, C.J. Taylor, D.W.L. Hukins, K. Hardinge, and M.L. Porter. Prosthetic
Hip Failure: Preliminary Findings of Retrospective Radiograph Image Analysis.
Enginnering in Medicine, 17:119125, 1988.
[54] J.M Kabo, J.S. Gebhard, G. Loren, and H.C. Amstutz. In Vivo Wear of Polyethylene
Acetabular Components. Journal of Bone and Joint Surgery, 75-B:254258, 1993.
[55] B. L. Kaptein, E. R. Valstar, B. C. Stoel, P. M. Rozing, and J. H. C. Reiber. A New
Model-based RSA Method Validated using CAD Models and Models from Reversed
Engineering. Journal of Biomechanics, 36:873882, 2003.
[56] J. Karrholm. Roentgen Stereophotogrammetry: An Overview of Orthopedic Applications. Acta Orthop Scand, 60:491503, 1989.
[57] Z. Knudsen, R.M. Hall, E. Berry, P. Siney, and B. Wroblewski. Assessment of Digital
Image Processing in the Measurement of Wear in Cemented THRs. Journal of Bone
and Joint Surgery, Supp II-B:135, 2001.
[58] R. Kosak, V. Antolic, V. Pavlovcic, V. Karlj-Iglic, I. Milosev, G. Vidmar, and
A. Iglic. Polyethylene Wear in Total Hip Prostheses: The Influence of Direction
of Linear Wear on Volumetric Wear Determined from Radiographic Data. Journal
of Skeletal Radiology, 32:679686, 2003.
[59] M. Krismer, R. Bauer, J. Tschupik, and P. Mayrhofer. EBRA: A Method to Measure
Migration of Acetabular Components. Journal of Biomechanics, 28:12251236, 1995.
[60] M. Krismer, B. Stockl, M. Fischer, R. Bauer, P. Mayrhofer, and M. Ogon. Early
Migration Predicts the Late Aseptic Failure of Hip Sockets. Journal of Bone and
Joint Surgery, 78-B:422426, 1996.
146
[61] D. Lawrie, M.R. Downing, G.P. Ashcroft, and P.H. Gibson. Incidence of ExtraOsseus Tantalum Beads in RSA of the Hip. Proceedings of the British Orthopaedic
Research Society, page 6, 2002.
[62] J. Livermore, D. Ilstrup, and B. Morrey. Effect of Femoral Size on Wear of the
Polyethylene Acetabular Components. Journal of Bone and Joint Surgery, 72A:518528, 1990.
[63] R.J. Looney, B. Boyd, S. Totterman, G.S. Seo, J. Tamez-Pana, D. Campbell,
L. Novotny, C. Olcott, J. Martell, F.A. Hayes, R.J. OKeefe, and E.M Schwarz.
Volumetric Computerized Tomogrophy as a Measurement of Prosthetic Acetabular
Osteolysis and Its Correlation With Wear. Arthritis Research, 4:5963, 2002.
[64] H. Malchau, P. Herberts, T. Eisler, G. Garellick, and P. Soderman. The Swedish
Total Hip Replacement Register. Journal of Bone and Joint Surgery, 84-A Suppl
2:220, 2002.
[65] J.M. Martell and S. Berdia. Determination of Polyethylene Wear in Total Hip Replacements with Use of Digital Radiographs. Journal of Bone and Joint Surgery,
79-A:16351641, 1997.
[66] H. McKellop, F.W. Shen, B. Lu, P. Campbell, and R. Salovey. Development of an
Extremely Wear-Resistant Ultra High Molecular Weight Polyethylene for Total Hip
Replacements. Journal of Orthopaedic Research, 17:157167, 1999.
[67] A.J. Murray, D.W. Carr and C.J. Bulstrode. Which Primary Total Hip Replacement? Journal of Bone and Joint Surgery, 77-B:520527, 1995.
[68] D. Nunn, M.A.R. Freeman, P.F. Hill, and S.J.W. Evans. The Measurement of
Migration of the Actebular Component of Hip Prostheses. Journal of Bone and
Joint Surgery, 71-B:629631, 1989.
[69] M.L. Olmstead. Canine Cemented Total Hip Replacements: State of the Art. Journal
of Small Animal Practice, 36:395399, 1995.
147
[70] S. Ordas, L. Boisrobet, and A.F. Frangi. Active Shape Models with Invariant Optimal Features (IOF-ASM) Application to Cardiac MRI Segmentation. Computers
in Cardiology, 30:633636, 2003.
[71] J.O. Otepka and C.S. Fraser. Accuracy Enhancement of Vision Metrology Through
Automatic Target Plane Determination. Proceedings of the International Society for
Photogrammetry and Remote Sensing 20th Congress, 35:873 879, 2004.
[72] E.T. Otterberg, D.L. Hauser, M.M. Siddiqui, C.R. Bragdon, and W.H. Harris. Radiographic Techniques to Assess in vivo Polyethylene Wear Following Total Hip
Arthroplasty. Harvard Orthopaedic Journal, 2, 2000.
[73] M. Pilu, A. Fitzgibbon, and R. Fisher. Ellipse-specific Direct Least-square Fitting.
IEEE International Conference on Image Processing, 1996.
[74] A.L. Redhead, A.C.W. Kotcheff, C.J. Taylor, M.L. Porter, and D.W.L. Hukins. An
Automated Method for Assessing Routine Radiographs of Patients With Total Hip
Replacements. Proceedings of the Institution of Mechanical Engineers, 211:145154,
1997.
[75] A.L. Redhead, A.C.W. Kotcheff, C.J. Taylor, M.L. Porter, and D.W.L. Hukins.
Quantitative Analysis of Total Hip Replacement Radiographs. Medical Image Understanding and Analysis, 1997.
[76] M. Rogers and J. Graham. Robust Active Shape Model Search. 7th European
Conference on Computer Vision, 4:517530, 2002.
[77] M. Rogers and J. Graham. Robust Active Shape Model Search for Medical Image
Analysis. Medical Image Analysis and Understanding, pages 8184, 2002.
[78] P.L. Rosin. Ellipse fitting by accumulating five-point fits. Pattern Recognition Letters, 14:661669, 1993.
[79] P.L. Rosin. Analyzing Error of Fit Functions for Ellipses. Pattern Recognition
Letters, 17:14611470, 1996.
148
[80] P.L. Rosin. Assessing Error of Fit Functions for Ellipses. Graphical Models and
Image Processing, 58(5):494502, 1996.
[81] P.L. Rosin. Further Five-Point Fit Ellipse Fitting. Graphical Models and Image
Processing, 61(5):245259, 1999.
[82] G. Roth and M.D. Levine. Extracting Geometric Primitives. CVGIP: Image Understanding, 58:122, 1993.
[83] R. Safaee-Rad, I. Tchoukanov, B. Benhabib, and K.C. Smith. Accurate Parameter Estimation of Quadratic Curves from Grey-Level Images. Computer Vision,
Graphics and Image Processing, 54:259274, 1991.
[84] M.E. Sarojak, W.A. Hoff, R.D. Komistek, and D.A. Dennis. An Interactive System
for Kinematics Analysis of Artificial Joint Implants. Proceedings of 36th Rocky
Mountain Bioengineering Symposium, 1999.
[85] M.E. Sarojak, W.A. Hoff, R.D. Komistek, and D.A. Dennis. Utilization of an Automated Model Fitting Process to Determine Kinematics of TKA. Proceedings of the
American Society of Biomechanics, 1999.
[86] G. Selvik. Roentgen Stereophotogrammetry: A Method for the Study of the Kinematics of the Skeletal System. Acta Orthopaedica Scandinavica, 232:151, 1989.
[87] S.M. Shaver, T.D. Brown, S.L. Hillis, and J.J. Callaghan. Digital Edge-Detection
Measurement of Polyethylene Wear after Total Hip Arthroplasty. Journal of Bone
and Joint Surgery, 79-A:693700, 1997.
[88] P.P. Smyth, C.J. Taylor, and J.E. Adams. Automatic Measurement of Vertebral
Shape Using Active Shape Models. Proceedings of the British Machine Vision Conference, pages 705714, 1996.
[89] C.J Sutherland, A.H. Wilde, L.S. Borden, and K.E. Marks. A Ten-Year Follow-up
of One Hundred Consecutive Muller Curver-Stem Total Hip Replacement Arthroplasties. Journal of Bone and Joint Surgery, 64-A:970982, 1982.
149
[90] G. Taubin. Estimation of Planar Curves, Surfaces, and Nonplanar Space Curves
Defined by Implicit Equations with Applications to Edge and Range Image Segmentation. IEEE Transactions of Pattern Analalysis and Machine Intelligence,
13(11):11151138, 1991.
[91] D. Terzopoulos and A. Witkin. Deformable models. IEEE Computer Graphics and
Applications, 8(6):4151, November 1988.
[92] E. Trucco and A. Verri. Introductory Techniques for 3-D Computer Vision. Prentice
Hall, 1998.
[93] E.R. Valstar, Jong F.W., Vrooman H.A., Rozing P.M., and Reiber J.H.C. ModelBased Roentgen Stereophotogrammetry of Orthopaedic Implants. Journal of Biomechanics, 34:715722, 2001.
[94] B. van Ginneken. Computer-Aided Diagnosis in Chest Radiography (PhD Thesis).
University Medical Centre Utrecht, 2001.
[95] B. van Ginneken, A.F. Frangi, J.J. Staal, B.M. ter Haar Romeny, and M.A.
Viergever. Active Shape Model segmentation with optimal features. IEEE Transactions on Medical Imaging, 21:924933, 2002.
[96] H.A. Vrooman, E.R. Valstar, G.J. Brand, D.R. Admiraal, P.M. Rozing, and J.H.
Reiber. Fast and Accurate Automated Measurements in Digitized Stereophotogrammetric radiographs. Journal of Biomechanics, 31:491498, 1998.
[97] S. Webb. The Physics of Medical Imaging. IOP Publishing Ltd., 1990.
[98] C.A. Wigderowitz, V. Sadaiyappan, D.M. McGurty, and D.I. Rowley. Computerized analysis of radiographs in Colles fractures. Clinical Orthopaedics and Related
Research, pages 217224, 2001.
[99] B.M. Wroblewski. Direction and Rate of Socket Wear in Charnley Low-Friction
Arthroplasty. Journal of Bone and Joint Surgery, 67-B:757761, 1985.
[100] B.M. Wroblewski. 15-21 Year Results of the Charnley Low-Friction Arthroplasty.
Clinical Orthopaedics and Related Research, 211:3035, 1986.
150
[101] P. Yao, G. Evans, and A. Calway. Face Tracking and Pose Estimation using Affine
Motion Parameters. Proceedings of the 12th Scandinavian Conference on Image
Analysis, 12:531536, 2001.
[102] Z. Zhang. Parameter Estimation Techniques: A Tutorial with Respect to Conic
Fitting. Image and Vision Computing, 15:5976, 1997.
[103] Y.H. Zhu, K.Y. Chiu, and W.M. Tang. Review Article: Polyethylene Wear and
Osteolysis in Total Hip Arthroplasty. Journal of Orthopaedic Surgery, 9(1):9199,
2001.
151

Thesis

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Thesis

Încărcat de

Drepturi de autor:

Formate disponibile

Automated Assessment of

Acetabular Wear From

A Thesis presented for the degree of

Computer Vision and Imaging Group

Aims and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Contents and contributions . . . . . . . . . . . . . . . . . . . . . . . . . . .

Structure of this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Total Hip Replacements . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Modelling Ellipses on Radiographs . . . . . . . . . . . . . . . . . . . 46

Data Points and Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Robust Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 60

Review of Ellipse Error Functions . . . . . . . . . . . . . . . . . . . . . . . . 63

Weighted Algebraic Distance . . . . . . . . . . . . . . . . . . . . . . 64

Angular Bisector of the Foci . . . . . . . . . . . . . . . . . . . . . . . 64

6 Eccentricity Error and Pose Estimation

CAD Model Based Pose Estimation . . . . . . . . . . . . . . . . . . 88

Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

7 Measuring Acetabular Wear

Uncorrected Wear Experiment . . . . . . . . . . . . . . . . . . . . . . . . . 99

Active Ellipses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

Measuring Wear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

Corrected Wear Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

Correction Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 118

A.1 Explicit Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

B.1 Overall System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

B.3 Contour Localization Subsystem . . . . . . . . . . . . . . . . . . . . . . . . 135

The femoral head and acetabular cup as they manifest on a radiograph

Region of interest of a healthy hip. 1) Acetabulum, the large cup-shaped

A total hip replacement. 1) Acetabular component, also called acetabular

Region of interest of a total hip replacement. 1) Acetabular rim wire that

Diagram of a cemented THR . . . . . . . . . . . . . . . . . . . . . . . . . . 10

Components of a cementless THR include (a) the acetabular cup consisting

Visualisation of the modes of failure. Image taken from Gruen et al.[41]. . . 14

Radiolucency of the femoral stem . . . . . . . . . . . . . . . . . . . . . . . . 15

2.10 Obtaining a THR radiograph through a simplified x-ray imaging system . . 19

Circle fitting in Polyware . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

Migration measured by Sutherland et al.s method. m1 is the caudal most

Migration measured by Nunns method. T) A line passing through the

An example radiograph with tantalum markers shown . . . . . . . . . . . . 36

The Active Ellipse Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 51

Template image and subsampled template . . . . . . . . . . . . . . . . . . . 55

Successful initialization by template and resultant convergence . . . . . . . 55

Region of interest and full view of radiograph where failure to localise

Region of interest of two radiographs with poor levels of contrast. (a)

The LMedS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Projecting both a circle and an ellipse from a circle . . . . . . . . . . . . . . 84

Correction of centres where (a) FFD=1300 and OFD=180 (b) FFD=1000

Correction of centres where (a) FFD=1300 and OFD=180 (b) FFD=1000

Distribution of wear over 4 years for annotator 1 . . . . . . . . . . . . . . . 109

7.10 Distribution of wear over 4 years for annotator 2 . . . . . . . . . . . . . . . 109

7.12 Distribution of wear over 4 years for annotator 4. Outlying measurements

List of nomenclature included in this thesis . . . . . . . . . . . . . . . . . . xix

Definitions for medical terms used in this thesis. . . . . . . . . . . . . . . . xxii

Modes of Cemented Femoral Stem Loosening . . . . . . . . . . . . . . . . . 14

Template Matching Success . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Eccentricity error values for given parameters. . . . . . . . . . . . . . . . . . 87

Observed Parameters of 2-D Wear . . . . . . . . . . . . . . . . . . . . . . . 113

Mean and standard deviation of uni-radiographic measurements performed

Mean and standard deviation of uni-radiographic measurements performed

Charles Stuart Kerrigan

Stephen McKenna, Ian Ricketts, Carlos Wigderowitz

Femoral head radius

Linear wear (mm)

Vector containing the 6 ellipse coefficients