Automatic morphometric cartilage quantification in the medial tibial plateau from MRI for osteoarthritis grading

Article (PDF Available)inOsteoarthritis and Cartilage 15(7):808-18 · July 2007with13 Reads
DOI: 10.1016/j.joca.2007.01.013 · Source: PubMed
Abstract
To evaluate whether a novel, fully automatic, morphometric cartilage quantification framework is suitable for assessing level of knee osteoarthritis (OA) in clinical trials. The population was designed with a normal population and groups with varying degree of OA of both sexes and at ages from 21 to 78. Posterior-anterior X-rays were acquired in semi-flexed, load-bearing position. The radiographic signs of OA were evaluated based on the Kellgren and Lawrence score (KL) and the joint space width (JSW) was measured. Turbo 3D T1 magnetic resonance imaging (MRI) scans were acquired with resolution 0.7x0.7x0.8mm(3) from a 0.18T scanner. The morphometric cartilage quantification from MRI resulted in volume, surface area, thickness and surface curvature for the medial tibial cartilage compartment. These quantifications were evaluated against JSW with respect to precision and ability to separate healthy subjects from OA subjects. The automatic, morphometric cartilage quantifications allowed fairly precise measurements with scan-rescan coefficient of variations (CVs) in the range from 3.4% to 6.3%. All quantifications, including JSW, allowed separation of the groups of healthy and OA subjects. However, for separation of the healthy from the borderline cases (KL 0 vs KL 1), only the Cartilage Curvature quantification allowed statistically significant separation (P<0.01). The novel morphometric framework shows promise for use in clinical trials. The ability of the Cartilage Curvature quantification to detect the early stages of OA and the effectiveness of the focal thickness Q10 measure are particularly noteworthy. Furthermore, these results may indirectly support that low-field MRI may be a low-cost option for clinical trials.
Automatic morphometric cartilage quantification in the medial tibial
plateau from MRI for osteoarthritis grading
E. B. Dam Ph.D.yzx*, J. Folkesson M.Sc.yk, P. C. Pettersen M.D.z and C. Christiansen M.D., Ph.D.z
y Image Group, IT University of Copenhagen, Denmark
z Center for Clinical and Basic Research, Ballerup, Denmark
x Nordic Bioscience Imaging, Herlev, Denmark
k Department of Computer Science, University of Copenhagen, Denmark
Summary
Objective: To evaluate whether a novel, fully automatic, morphometric cartilage quantification framework is suitable for assessing level of knee
osteoarthritis (OA) in clinical trials.
Method: The population was designed with a normal population and groups with varying degree of OA of both sexes and at ages from 21 to
78. Posterioreanterior X-rays were acquired in semi-flexed, load-bearing position. The radiographic signs of OA were evaluated based on the
Kellgren and Lawrence score (KL) and the joint space width (JSW) was measured. Turbo 3D T1 magnetic resonance imaging (MRI) scans
were acquired with resolution 0.7 0.7 0.8 mm
3
from a 0.18 T scanner. The morphometric cartilage quantification from MRI resulted in vol-
ume, surface area, thickness and surface curvature for the medial tibial cartilage compartment. These quantifications were evaluated against
JSW with respect to precision and ability to separate healthy subjects from OA subjects.
Results: The automatic, morphometric cartilage quantifications allowed fairly precise measurements with scanerescan coefficient of variations
(CVs) in the range from 3.4% to 6.3%. All quantifications, including JSW, allowed separation of the groups of healthy and OA subjects. How-
ever, for separation of the healthy from the borderline cases (KL 0 vs KL 1), only the Cartilage Curvature quantification allowed statistically
significant separation (P < 0.01).
Conclusion: The novel morphometric framework shows promise for use in clinical trials. The ability of the Cartilage Curvature quantification to
detect the early stages of OA and the effectiveness of the focal thickness Q10 measure are particularly noteworthy. Furthermore, these results
may indirectly support that low-field MRI may be a low-cost option for clinical trials.
ª 2007 Osteoarthritis Research Society International. Published by Elsevier Ltd. All rights reserved.
Key words: Cartilage, Quantification, Low-field MRI, Volume, Area, Thickness, Curvature.
Introduction
Osteoarthritis (OA) is affecting the daily lives of the majority
of the older part of the world population d for some by mi-
nor morning stiffness and for others to the degree of caus-
ing severe pain, joint swelling, reduced range of motion, and
disability
1
. Thereby, it is a severe cause for reduced quality
of life. The total economic burden of arthritis is estimated at
1e2.5% of the gross national product of Western countries;
OA accounts for the major share
2
.
Furthermore, even though promising new treatment pos-
sibilities are arising, a major, thoroughly documented break-
through in effective treatment of OA beyond symptom
control is lacking. One limiting factor in the development
and evaluation of new treatments is the effectiveness of
the methods for quantification of disease progression in
clinical trials evaluating the effect of potential disease-mod-
ifying osteoarthritis drugs (DMOADs) such as calcitonin
3
.
Accuracy (/correctness) and precision (/reproducibility) of
the quantification methods are essential together with the
ability to quantify current level of the disease as well as to
monitor the actual progression of the disease. These factors
affect both the number of test subjects needed in a clinical
study and the required duration of the study. In addition to
focusing on accuracy and precision, we wish to stress the
value of automating the quantification methods d in line
with the recommendations from the US Food and Drug Ad-
ministration: ‘‘Precision is the goal’’
4
. Fully automatic (typi-
cally computer based) quantification methods by definition
eliminate intra- and inter-observer variation and thereby po-
tentially allow better precision. Furthermore, for studies
based on medical imaging data (X-ray, magnetic resonance
imaging (MRI), computed tomography (CT), etc.), the load
on the radiologists is potentially overwhelming d and in-
creasingly so when three dimensional (3D) morphometric
measures are desired. Advanced morphometric measures
such as curvature require fully segmented structures with
anatomical correspondence defined. Therefore, computer-
based methods can not only relieve the radiologists but
also allow quantification measures that would otherwise
be infeasible in large-scale studies.
The currently accepted standard for monitoring OA pro-
gression in clinical trials of DMOADs is quantification of
the joint space width (JSW) measured in X-rays
5
. However,
we focus on quantification of articular knee cartilage from
MRI which offers advantages compared to traditional
X-ray based OA monitoring. First, the cartilage is visible and
*Address correspondence and reprint requests to: Erik B. Dam,
Nordic Bioscience Imaging, Herlev Hovedgade 207, 2730 Herlev,
Denmark. Tel: þ45-4454-7777; Fax: þ45-4454-8888; E-mail:
erikdam@nordicbioscience.com, http://www.diku.dk/werikdam
Received 28 June 2006; revision accepted 16 January 2007.
OsteoArthritis and Cartilage (2007) 15, 808e818
ª 2007 Osteoarthritis Research Society International. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.joca.2007.01.013
International
Cartilage
Repair
Society
808
MRI allows visualization of both cartilage boundaries and
cartilage structure
6
. Second, by using 3D scans morpho-
metric analysis is possible.
Several computer-based, semi-automatic methods for
quantifying cartilage have been published. Among the
most prominent segmentation methods are the semi-auto-
matic slice-wise B-spline-based method
7
, the watershed-
based method
8
, and the region-growing method
9
. These
methods also form the basis for evaluation of morphometric
measures
10e12
. The need for advanced morphometric
quantification is supported by the slow evolution of cartilage
loss as reported in Gandy et al.
13
where no cartilage loss
was observed over 3 years for OA patients. More details
on related methods are given in the discussion.
We introduce a framework for morphometric cartilage
quantification using a novel, fully automatic computer-
based method. Strictly speaking morphometry means
quantification of shape d we use the term broadly for
both simple measures as volume and shape-related
measures such as thickness and curvature. The framework
allows quantification of the medial tibial cartilage compart-
ment in terms of volume, surface area, thickness, and
surface curvature. Volume and area quantify the overall
cartilage loss, whereas thickness is a natural measure for
quantifying focal cartilage breakdown. Curvature is a mea-
sure of the bending of the cartilage surface and is related
to the joint biomechanics and thereby potential cartilage
breakdown rather than to quantification of past cartilage loss.
The aim of this study was to evaluate whether our quan-
tification framework was suitable for quantifying the degree
of OA with high precision d and in particular whether the
early stages were detectable.
Subjects and method
POPULATION
A population of test subjects was prospectively selected.
The subjects were randomly selected such that the popula-
tion had an even distribution between sexes and across
ages and such that the number of healthy and subjects
with varying degree of OA symptoms were approximately
equal. The majority of the subjects were invited from ad-
dress lists, but the population also contained volunteers
with known knee problems (and therefore likely to have
OA). Subjects with previous knee joint replacement, inflam-
matory arthritis or presenting any contraindication for MRI
examination were excluded prior to the study.
A total of 144 test subject knees were imaged for the
study. Among these, five were discarded due to insufficient
image quality in either radiograph or MRI scan d leaving
139 test subjects knees in the study. This group was ran-
domly divided into a training set of 25 knees (used for train-
ing of the automatic methods) and an evaluation set of the
remaining 114 knees. Both left and right knees were
included.
In the evaluation set, 51 knees were healthy and 63 had
some degree of OA (a similar distribution was present in the
training set). A subset of 31 from the 114 knees were re-
scanned a week after the first scan. Further details are
listed in Table I.
All participants signed approved information consent and
the study was carried out in accordance with the principles
of the Helsinki Declaration II and European Guidelines for
Good Clinical Practice. The study protocol was approved
by the local Ethical Committee.
In this study, we focused on the medial compartment of
the tibial cartilage layer since the correlation between deg-
radation and clinical symptoms is predominant in the medial
compartments
14
and in particular in the tibial part
15
. Further-
more, other studies show that the cartilage loss is highly
correlated between tibial and femoral cartilage and that
the measurements of the tibial cartilage are more reproduc-
ible
16
. Following the nomenclature in Eckstein et al.
17
, this
is the MT anatomical region.
X-RAY PROTOCOL
For each test subject, digital X-rays of both knees were
acquired. The subject was positioned standing in a weight
bearing position with the knees slightly flexed and the feet
rotated externally. In order to optimize reproducibility, we
used the SynaFlex (developed by Synarc) to fix the orienta-
tion of the feet and flexing of the knees
18
.
Focus film distance was 1.0 m, and the tube was angu-
lated in 10
(the metatarsophalangeal (MTP) view modified
for fixed angle
19
). Radiographs were acquired in the poste-
rioreanterior (PA) position, while the central beam was dis-
played directly to the mid point of the line passing through
both popliteal regions. Radiographs of both knees were ac-
quired simultaneously.
MRI PROTOCOL
We acquired knee MRI using an Esaote C-Span 0.18 T
scanner dedicated for imaging of the extremities. We
used the same knee coil for all subjects. The sequence
was a sagittal Turbo 3D T1 (flip angle 40
, repetition time
50 ms, echo time 16 ms, number of acquisition averages 1)
with scan time around 10 min. The resolution was
Table I
The evaluation population characteristics with regards to count,
gender, age, BMI and degree of OA given by the KL index
Evaluation set Scanerescan set
N 114 31 (2)
Gender 54% female 55% female
Age 21e78, mean 55 26e75, mean 61
BMI 20e38, mean 27 20e33, mean 25
KL: N on 0e4 51, 28, 14, 21, 0 11, 12, 2, 5, 0
Fig. 1. A sagittal slice from the Turbo 3D T1 MRI of a 67-year-old
male subject with KL index 3. The contours are the manual outlines
of femoral and tibial cartilage performed by a radiologist.
809
Osteoarthritis and Cartilage Vol. 15, No. 7
0.70 mm 0.70 mm in each slice with a slice thickness be-
tween 0.70 mm and 0.93 mm but typically at 0.78 mm (with
no gap between slices). Each slice is 256 256 pixels and
a scan has around 110 slices depending on the size of the
knee. See Fig. 1 for an illustration of a scan. We used a 3D
sequence with near-isotropic voxels since this is well suited
for cartilage quantification
20
and for 3D modeling in general.
The test subjects were imaged while lying with no loading
prior or during the scanning except for walking to the
scanner.
The articular cartilage in the tibial and femoral medial
compartment was manually segmented by slice-wise delin-
eation by a radiologist (see Fig. 1). These manual segmen-
tations were used for training of the automatic methods and
for comparative evaluation of the volume quantification.
RADIOGRAPH ANALYSIS
From the X-rays, the radiologist determined the Kellgren
and Lawrence (KL) score
21
on the scale from 0 (healthy)
to 4 (severe OA). The score is based on a qualitative eval-
uation of presence of osteophytes, joint gap narrowing, and
in the severe cases sclerosis of the subchondral bone.
The JSW was measured from the X-rays by the radiolo-
gist by manually marking the narrowest gap between the fe-
mur and the tibia in the medial compartment.
The radiologist also marked the most medial and lateral
points on the tibial plateau (not including possible osteo-
phytes). These points define the tibial plateau width which
is a simple measure of the size of the knee that we used
for normalization of the quantifications.
AUTOMATIC CARTILAGE SEGMENTATION FROM MRI
The first step in the fully automatic, morphometric quanti-
fication framework is voxel classification based on super-
vised learning
22
. This scheme is augmented by a virtual
normalization of the placement of the knee in the scanner
23
.
The voxel classification describes each voxel by features
quantifying local scan texture properties such as position,
intensity, edgeness, flatness, orientation, and ridgeness
(based on Gaussian derivatives up to third order at multiple
scales). These feature vectors from the voxels from the 25
scans in the training set form a database of examples of tib-
ial cartilage, femoral cartilage and background voxels. The
automatic segmentation on a new scan is performed by
classifying each voxel using two k Nearest Neighbor classi-
fiers that first locate tibial and femoral cartilage separately
followed by a voting that ensures non-overlapping cartilage
layers. Classification outliers are removed by selecting the
largest connected component. During the automatic steps,
right knees are mirrored around the center sagittal plane
in order to manage all knees in the same framework.
The fully automatic segmentation required 10 min of com-
putation (on a standard 2.8 GHz desktop computer) using
an optimized algorithm for voxel classification
24
. The seg-
mentation mean accuracy was previously evaluated against
segmentations performed by manual outlining by a radiolo-
gist on the same evaluation data set used in this study. The
mean accuracy was given by sensitivity 83.9% and specific-
ity 99.9% (sensitivity is the percentage of the cartilage out-
lined by the radiologist captured by the method, specificity
is the percentage of background not captured). This was
comparable to a testeretest evaluation where the radiolo-
gist performed repeated manual segmentations on the
same set
23
. An example segmentation is illustrated in
Fig. 2A.
The morphological quantification steps in the following
sections were performed on the result of the automatic
voxel classification segmentations.
TIBIAL CARTILAGE SHAPE MODEL
A voxel classification alone does not allow advanced
morphometric quantification. In order to do that properly,
a parametric representation of the cartilage surface is
needed. Therefore, we fit a shape model to the result of
the automatic segmentation d as illustrated in Fig. 2.
We used a shape model, the m-rep
25
, and deformed it
such that the model boundaries were aligned with the tran-
sition between cartilage and background. This was done by
optimizing the parameters for the shape model in a Bayes-
ian framework where the statistical shape model trained
from the 25 training shapes ensured a plausible cartilage
shape while the model boundary approaches the cartilage
boundary driven by a boundary distance transformation per-
formed on the result of the automatic segmentation
26
. The
statistical shape model was trained on mainly healthy sub-
jects such that the preferred shapes will be those of healthy
cartilage layers. However, the deformation of the model al-
lows regions with zero thickness.
The deformation of the shape model was fully automatic
and took 10 min of computation. The quantification steps
below were also fully automatic and required no significant
computation time.
QUANTIFICATION OF CARTILAGE VOLUME
The Cartilage Volume was computed directly from the au-
tomatic segmentation by counting the number of voxels
Fig. 2. (A) The automatic segmentation of the medial compartment of the tibial cartilage from the scan in Fig. 1. (B) The shape model fitted to
the segmentation. The model is shown as a wire-frame with the internal skeleton that defines the model coordinate system in red. (C) The
model shown as a surface model.
810
E. B. Dam et al.: Automatic morphometric cartilage quantification
classified as belonging to the medial tibial cartilage layer.
For comparison, we also computed the cartilage volume
from the manual segmentations performed by a radiologist.
Following the nomenclature from Eckstein et al.
17
, this
measure is denoted MT.VC. We denote the volumes from
the manual segmentations MT.VC.man.
QUANTIFICATION OF CARTILAGE AREA
Like cartilage volume, the cartilage area also aims at
quantifying the overall cartilage loss. However, a combina-
tion of the two measurements can give a rough indication of
whether the cartilage loss is an overall thinning (loss of vol-
ume, little loss of area) or a lesion-based loss with gaps and
holes in the cartilage layer (loss of volume, large loss of
area).
The Cartilage Area was computed from the surface repre-
sentation given by the cartilage shape model illustrated in
Fig. 2 (as the summation of the areas of the generated sur-
face elements).
Following the nomenclature from Eckstein et al.
17
, this
measure is denoted MT.AC. Note that the entire surface
area of the cartilage layer is included d both the area of
the cartilage/bone interface and the area of the superior sur-
face of the layer.
QUANTIFICATION OF CARTILAGE THICKNESS
Cartilage thickness is also intended to quantify cartilage
loss. However, a thickness map directly allows detection
of local variation and thereby analysis of whether the
cartilage loss is caused by overall thinning or by focal
lesions.
The m-rep is a medial shape model (medial in the math-
ematical sense). This means that the basic atoms of the
model are in the center of the objects and have the radius
as the central attribute. Thereby, the choice of shape model
makes it fairly simple to extract a thickness map across the
cartilage layer d as illustrated in Fig. 3.
In the remainder of this paper, we use the term Cartilage
Thickness as the mean of the thickness map excluding the
edges of the cartilage layer at the periphery of the tibial pla-
teau (around the crest of the shape model). The crest was
excluded since the precision in both position and thickness
quantification can be relatively low
27
.
In order to investigate whether the cartilage loss is mainly
an overall thinning or more present as focal lesions, we also
measured the thickness of the thinnest part of the cartilage
layer (disregarding the location of this thinnest part). We
use the term Cartilage Thickness Q10 for the 10% quantile
of the measurements from the cartilage thickness map.
Following Eckstein et al.
17
, the Cartilage Thickness is
MT.ThCtAB (again also covering the denuded area of the
bone) and Cartilage Thickness Q10 is MT.ThCtAB.Q10.
QUANTIFICATION OF CARTILAGE CURVATURE
Analogous to the thickness map, we computed a surface
curvature map for the cartilage layer. The curvature was lo-
cally approximated by the change in cartilage surface nor-
mals in a small neighborhood. The surface normals were
given by the shape model. For more details, see Folkesson
et al.
28
. The anatomical meaning of a curvature quantifica-
tion depends on the scale at which the measurement is per-
formed. At fine scale, the surface curvature is a measure of
the smoothness/roughness; at coarse scale, the curvature
is related to the overall shape and thereby the joint congru-
ity. Our method is at a finer scale than previous methods
(most prominently
10
) but is still focused at the overall shape.
The curvature map is illustrated in Fig. 4. Since the curva-
ture of the model is dominant at the edges of the cartilage
layer (the crest of the shape model), we defined a region
of interest that includes the load-bearing region of the carti-
lage layer but excludes the crest.
Since both the neighborhood for approximating the local
curvature as well as the region of interest were defined in
terms of the coordinate system of the shape model (illus-
trated in Fig. 2), their sizes varied slightly between scans.
1.2
1.4
1.6
1.8
2.0
2.2
2.4
2.6
2.8
3.0
AB
Thickness Map
Medial
Knee Center AnteriorPosterior
Fig. 3. (A) Map of the local thickness across the surface of the cartilage layer shape model from Fig. 2. Bright areas are thicker. (B) Using the
coordinate system given by the shape model, this map can also be illustrated in a regular grid. The unit is mm.
811
Osteoarthritis and Cartilage Vol. 15, No. 7
On average, the neighborhood was 3 by 3 mm (standard
deviation 0.4 mm), and the region of interest was 16 by
24 mm (longest in the anteriore posterior direction) with
standard deviation 2 by 3 mm.
In the remainder of this paper, we use the term Cartilage
Curvature for the mean of the local curvature approximation
across the load-bearing region of the cartilage layer defined
above. Extending on the nomenclature principles from
Eckstein et al.
17
, the Cartilage Curvature is denoted MT.CuC.
KNEE SIZE NORMALIZATION
In order to prevent the knee size to confound the results,
we normalized the measures by the width of the tibial pla-
teau. Since osteophytes were excluded when measuring
the tibial width, this normalization was intended to account
only for knee size differences (related to age, sex or other
growth factors) and not any effects related to bone remod-
eling or presence of OA.
The normalization was done by the tibial width to the
power of the length unit such that each measure becomes
unit-less and thereby scale-invariant (e.g., volume was nor-
malized by the plateau width cubed).
Since curvature has unit mm
1
it was normalized by the
inverse of the tibial width. This scale-invariance is anatom-
ically meaningful since we measured the overall bending of
the shape and the local bending would increase for smaller
shapes without normalization. If the surface curvature was
defined at a smaller scale where the surface smoothness
could be measured (such as in Folkesson et al.
28
), this
scale-invariance would possibly not be meaningful since
the surface smoothness is likely to be related to the size
of the cartilage building blocks (collagen, aggrecan, .)
and not the size of the knee.
EVALUATION AND STATISTICS
We evaluated the precision/reproducibility of our quantifi-
cations by comparing the measures on pairs of scans of the
same knee acquired with a week in between. We measured
this in terms of linear correlation coefficient (COR), mean
relative absolute difference (RAD), and mean coefficient
of variation (CV) between the pairs of values.
We evaluated whether the quantifications are suitable for
monitoring OA progression (defined by the KL score) by
checking if they allowed separation of healthy and OA
knees. This was done by performing an unpaired t test on
the values d the resulting P value estimates the probability
for the two groups not being different. So low P values, typ-
ically defined to be below 0.05, show a statistically signifi-
cant separation of the two groups. Since we are
introducing a handful of quantifications, there is a possibility
that the resulting P values will, by chance, be lower for a sin-
gle of them. The Bonferroni correction is a conservative
means for avoiding this. Since we were testing five
measures, we could therefore conservatively estimate our
P values to a factor of 5 higher than what they were. Instead,
we simply chose to use 0.01 as our level of significance.
Also, in the same manner we evaluated whether the
quantifications allowed separation of the healthy from the
mild OA cases (defined as KL score 1 or 2), and finally eval-
uated whether separation of the healthy from the borderline
OA cases (defined as KL 1) was possible.
Results
The graph in Fig. 5(A) illustrates the ability of the JSW
quantification to separate healthy from OA subjects and
then the various degrees of OA as defined by KL. For
A
B
C
Curvature Map (Load)
Medial
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.10
Knee Center AnteriorPosterior
Fig. 4. (A) Map of the local curvature across the surface of the cartilage layer shape model from Fig. 2. (B) Since the curvature is dominant
around the crest of the shape, we define a region of interest surrounding the load-bearing area. The intensities are rescaled to reveal the areas
with large surface bending as bright areas. (C) Like the thickness map in Fig. 3, we extract the thickness map in a regular grid defined by the
shape model. The unit is mm
1
.
812
E. B. Dam et al.: Automatic morphometric cartilage quantification
perspective, the performance of the cartilage volume from
the radiologist manual segmentations is illustrated in
Fig. 5(B).
For the automatic Cartilage Volume quantification, both
the scanerescan precision (A) and the ability to separate
levels of OA (B) are illustrated in Fig. 6.
The ability to separate levels of OA are illustrated analo-
gously for the automatic quantifications of Cartilage Area,
Cartilage Thickness, Cartilage Thickness Q10, and Carti-
lage Curvature quantifications in Fig. 7.
The results for precision are summarized in Table II. All
results were obtained by comparing measures from the
31 knees in the scanerescan collection. The precision is
given as the mean RAD, the mean CV, and the linear
COR. For N knees with a quantification Q of first scan
Q
a
i
and quantification on the rescan a week later Q
b
i
(where
i ¼ 1.N ), we defined:
RAD ¼
1
N
X
N
i¼1
Q
a
i
Q
b
i
mean
Q
a
i
; Q
b
i
and
CV ¼
1
N
X
N
i¼1
std
Q
a
i
; Q
b
i
mean
Q
a
i
; Q
b
i
The results for the ability to separate healthy from OA
knees are summarized in Table III with the P values from
the three comparisons of the healthy group vs all OA (KL
above 0), mild OA (KL 1 or 2), and borderline OA (KL 1).
0>0 01234
1.0
2.0
3.0
4.0
5.0
**
**
Kell
g
ren & Lawrence Index
Joint Width Space [mm]
Mean and SEM
0>0 01234
1300
1400
1500
1600
1700
1800
1900
2000
**
Kell
g
ren & Lawrence Index
Cartilage Volume (manual) [mm
3
]
Mean and SEM
AB
Fig. 5. (A) JSW and (B) cartilage volume from manual segmentation (MT.VC.man). For each, the mean measurements are shown (with bars
illustrating the standard error of the mean) for the groups healthy and OA and then, to the right of the dotted line, for each KL score. The levels
where statistically significant separation is possible are marked by stars.
1000 2000 3000
1000
2000
3000
Volume [mm
3
]
(
first scan
)
Volume [mm
3
]
(second scan)
Scan−Rescan Precision
0 >0 0 1 2 3 4
1500
2000
2500
*
*
Kell
g
ren & Lawrence Index
Volume [mm
3
]
Mean and SEM
AB
Fig. 6. Cartilage Volume (MT.VC). (A) The scanerescan precision is illustrated. (B) The ability to separate levels of OA defined by KL.
813
Osteoarthritis and Cartilage Vol. 15, No. 7
Discussion
Automated, computer-based methods can save time for
expert readers in clinical studies and potentially provide
high precision due to the elimination of inter/intra-reader
variation. Equally interestingly, such methods can allow
quantification that would otherwise be infeasible such as
a thickness map or a mean curvature estimation. In our dis-
cussion we focus on measurement precision and the ability
to quantify the level of OA with a particular interest in the
early stages.
GOLDEN STANDARD FOR LEVEL OF OA
We chose to accept the KL score as the ground truth for
OA progression. Admittedly, the KL score is a somewhat
simplified quantification with a surprisingly large variation
in the scores determined by different observers
29
. Further-
more, it could be argued that a whole-organ model (such
as the WORMS
30
) would be superior. We still chose the
KL score due to its relative simplicity and since it is
generally accepted. Since the KL score has a relatively
large focus on bone, our cartilage quantifications may be
even better for quantifying level of OA with a better defini-
tion of OA.
Table II
Quantification precision. For each quantification measure, the pre-
cision is given as mean RAD, mean CV, and linear COR. The eval-
uationisperformedon31scanerescanpairs.ForCV,alsothe
standard deviation and the upper limit for the 95% confidence inter-
val of the values are given parenthesized
Quantification RAD (%) CV (%) COR
JSW (manual, X-ray) 4.0 2.8 (2.7, 8.1) 0.99
Volume (MT.VC.man)
(manual MRI)
10.3 7.3 (4.9, 16.9) 0.90
Volume (MT.VC) 5.9 4.1 (4.7, 13.3) 0.91
Area (MT.AC) 5.5 3.9 (3.9, 11.6) 0.89
Thickness (MT.ThCtAB) 4.7 3.4 (3.6, 10.4) 0.83
Thickness Q10 (MT.ThCtAB.Q10) 4.7 3.3 (2.9, 8.7) 0.88
Curvature (MT.CuC) 9.0 6.3 (5.6, 17.3) 0.64
0 >0 0 1 2 3 4
1700
1800
1900
2000
2100
2200
2300
2400
2500
***
Kellgren & Lawrence Index
Area [mm
2
]
Mean and SEM
0 >0 0 1 2 3 4
1.9
2.0
2.1
2.2
2.3
2.4
2.5
*
*
Kellgren & Lawrence Index
Thickness [mm]
Mean and SEM
0>0 01234
1.2
1.3
1.4
1.5
1.6
1.7
***
**
Kell
g
ren & Lawrence Index
Thickness Q10 [mm]
Mean and SEM
0 >0 0 1 2 3 4
0.035
0.040
0.045
0.050
0.055
0.060
0.065
0.070
***
**
*
Kell
g
ren & Lawrence Index
Curvature [mm
−1
]
Mean and SEM
A
B
CD
Fig. 7. The ability to separate levels of OA for (A) Cartilage Area (MT.AC), (B) Cartilage Thickness (MT.ThCtAB), (C) Cartilage Thickness Q10
(MT.ThCtAB.Q10), and (D) Cartilage Curvature (MT.CuC).
814
E. B. Dam et al.: Automatic morphometric cartilage quantification
ACCURACY VALIDATION
We would like to evaluate the accuracy of the measure-
ments. Unfortunately, direct evaluation of the accuracy of
the morphometric measures is problematic on human test
subjects in vivo. Much effort has been put into validating
cartilage quantification from MRI, for instance using cadav-
eric joints or animal joints
12,31e33
or from synthetic
models
34
. Alternatively, joints acquired after joint replace-
ment surgery can be used. These validations are admirable,
but problematic to pursue for larger populations especially
for healthy test subjects.
Furthermore, for some morphometric measures, such as
area and curvature, even putting ethical and practical con-
siderations aside, it is not obvious how to measure ground
truth values. Therefore, we focus on evaluating the ability of
the measures to differentiate various degrees of OA accord-
ing to the KL score. Thereby, we also indirectly evaluate the
accuracy of the quantification methods.
RELATED WORK AND PRECISION
Several methods exist for segmentation and quantifica-
tion of articular knee cartilage from MRI. To the best of
our knowledge, no other fully automatic method has been
evaluated and published. In general, the MRI based
methods are more often evaluated for segmentation accu-
racy and precision rather than for morphometric quantifica-
tion and ability to differentiate degrees of OA according to
KL. Some key papers on cartilage quantification are listed
in Table IV (for a review, see Eckstein et al.
35
).
In Ding et al.
36
they show that cartilage volume is signif-
icantly related to presence of cartilage defects in a study in-
cluding 372 test subjects. A relationship between cartilage
volume and progression of OA is also reported in Cicuttini
et al.
16
, Jones et al.
37
, and Amin et al.
38
.
Relatively little research has been done on the connec-
tion between the shape of the cartilage layer and disease
progression. The biomechanics of the joint both affect and
are affected by the shape. There is also a dual connection
with the shape and the structure of the underlying bone.
The shape of the cartilage affects the strain on the bone
and the shape of the bone caused by bone remodeling is
directly reflected in the shape of the cartilage. However,
both Hohe et al.
10
and Terukina et al.
39
introduce methods
for quantifying the cartilage curvature. In Hohe et al.
10
they
report measures of curvature for all knee cartilage compart-
ments, with a standard deviation of 4.7 m
1
on repeated
measurements of mean curvature of the medial tibial carti-
lage where the mean value for the population is 29.6 m
1
.
However, to the best of our knowledge, no studies prior to
this have been published evaluating if curvature is a signifi-
cant OA disease marker.
The evaluation of our quantification framework showed
precision comparable to the results from the literature
(see Table IV). Specifically, for Cartilage Thickness, we
had a CV of 3.4% to be compared with the values
4.6%
40
,5%
41
, 6.6%
12
, and 2.3%
42
. For Cartilage Area,
we had a CV of 3.9% to be compared to 2.5%
10
and
2.9%
42
. It is difficult to compare the values for Cartilage Cur-
vature directly due to difference in the scale at which the
cartilage curvature was measured d this is evident in the
mean values for the populations: 29.6 m
1
in Koo et al.
12
compared to our values around 45 m
1
(or 0.045 mm
1
).
However, their reported standard deviation and mean
values give an approximate CV of 16%. Our scanerescan
measurements had a mean CV of 6.3%.
Our precision numbers seem comparable to the existing
methods. However, it is not trivial to compare our precision
numbers directly against the values from the literature.
Firstly, different populations with differing ages, levels of
OA, body mass index (BMI) and other characteristics influ-
ence the results. Secondly, all the studies mentioned above
base their measurements on sequences from high-field
scanners.
LOW-FIELD VS HIGH-FIELD
A central choice is the use of low-field MRI. Most recent
work is focused on high-field MRI using specialized se-
quences dedicated to cartilage quantification from MRI
(for a review, see Peterfy et al.
43
). However, we chose to
investigate whether low-field MRI could be suitable for
use in clinical trials. Low-field scanners are much cheaper
and easier to install and maintain since no cooling of
super-conductors is needed. The scanner used in this study
has a permanent magnet and only needs a standard
power outlet to run. Thereby, the use of low-field scanners
can potentially lower costs in clinical trials considerably.
Recently, fat suppression sequences have also been
demonstrated on a low-field 0.35 T scanner
44
. Thereby,
the scan quality of low-field MRI for cartilage quantification
could possibly improve further in the future.
Even if the precision of our methods seem comparable to
many methods based on high-field MRI, we can’t quite com-
pete with the precision from manual outlining in 3 T
DESSwe scans
42
. However, the big differences in method-
ology do not allow a proper conclusion on whether low-field
MRI in itself currently allows comparable performance. Fur-
ther research with a direct comparison using the same pop-
ulation is needed for that.
It should be noted that our methodological framework is
independent of the choice of low-field MRI. All the steps
just need a training collection of annotated training scans
like the one we used in this work in order to be applied to
other sequences or other scanners.
PRECISION AND INTERACTION
Any fully automated computer-based method will to some
degree fail in some cases for such a challenging task as
cartilage segmentation and quantification. This is evident
in Fig. 6 where a single among the 31 knees is clearly off
the diagonal with automatic volume quantifications of
2274 mm
3
and 3172 mm
3
in the scanerescan pair.
Table III
Ability to separate groups of healthy from OA test subjects. In the
test population there are 51 healthy (KL 0), 61 OA (KL above 0),
42 mild OA (KL 1 or 2), and 28 borde rline OA (KL 1 ) subjects.
The table shows P values from an unpaired t test for each quanti-
fication method and for each group comparison
Quantification Healthy vs
all OA
Healthy vs
mild OA
Healthy vs
borderline
JSW (manual, X-ray) 0.005 0.6 0.9
Volume (MT.VC.man)
(manual, MRI)
0.0016 0.04 0.16
Volume (MT.VC) 0.017 0.3 0.4
Area (MT.AC) 0.00069 0.052 0.27
Thickness (MT.ThCtAB) 0.013 0.3 0.2
Thickness Q10
(MT.ThCtAB.Q10)
0.00015 0.051 0.3
Curvature (MT.CuC) 0.0000044 0.00034 0.0052
815
Osteoarthritis and Cartilage Vol. 15, No. 7
Apart from methodological shortcomings, an automatic
method will also be limited by the amount of training exam-
ples and in addition have an implicit bias toward the radiol-
ogist who performed the manual segmentations introduced
in the training.
In a diagnostic setting this would be unacceptable. Then
an interactive correction of the segmentation would be cru-
cial d like the one based on an extended watershed trans-
formation in Dam et al.
45
. In clinical studies, the advantages
of automation may outweigh a minor loss of precision
caused by few outliers. And the bias toward an expert radi-
ologist is what centers with multiple readers aim for through
careful training of the readers.
SEPARATION OF HEALTHY AND OA AND DETECTION
OF EARLY OA
The population was designed with a large normal popula-
tion with little or no OA symptoms and a group of subjects
with varying degree of OA. This composition allowed not
only investigation of the ability to separate healthy from
OA subjects, but also allowed investigation of the ability to
detect the early stages of OA.
Since the joint gap inspected from the X-ray is directly
influencing the KL scores, it is reasonable to expect a close
correlation between the JSW and the KL score. However,
the more automatic, morphometric quantifications were
comparable or superior in the ability to separate the healthy
(KL 0) from the knees with OA (KL above 0). This was the
case even if some of the more complex, automated mea-
sures were not quite as precise as the simple JSW mea-
surement. This clearly indicates that the morphometric
measures of cartilage shape actually do capture the dis-
ease level better than the joint gap measurement.
Two of the quantifications stand out as being particularly
promising. First, the Cartilage Thickness Q10 measure was
both precise (CV 3.3%) and could separate healthy from
OA with high statistical significance (P < 0.001). However,
Thickness Q10 allowed only weak discrimination between
healthy and mild OA and no discrimination between healthy
and borderline OA. Second, the Cartilage Curvature mea-
sure could distinguish the group of healthy knees not only
from the group of healthy (P < 0.0001), but also from the
group of knees with mild OA (P < 0.001) as well as the
group with borderline OA with statistical significance
(P < 0.01).
OA ETIOLOGY
Even if many factors regarding the prevalence and pro-
gression of OA have been shown, there is still no clear un-
derstanding of what the causes of OA are. The results
obtained in this study allow some speculation in this
direction.
One central question is whether cartilage breakdown is
mainly a global, gradual thinning (possibly indicating a sys-
temic cause) or whether it is a focal thinning (indicating
breakdown caused by local lesions). The fact that the Thick-
ness Q10 measure based on the 10% quantile measure
can separate healthy from OA subjects much stronger
than the mean Thickness measure indicates that focal le-
sions are very central in the cartilage breakdown.
Secondly, since the Curvature measure can separate
healthy (KL 0) from borderline OA (KL 1) subjects, there is
a strong indication that the early onset of OA is related to
joint congruity and mechanical stress in the joint. Future re-
search will show whether the curvature measure is related
to knee alignment, exercise or other factors and thereby
potentially guide the research into prevention of OA.
MORPHOMETRIC CARTILAGE QUANTIFICATION IN CLINICAL
TRIALS
The framework for morphometric cartilage quantification
presented here provided quite precise measurements that
allowed separation of the group of healthy subjects from
both mild and borderline OA (using Curvature) as well as
clear separation of healthy from OA overall (several quanti-
fications, but particularly strong using Thickness Q10). We
therefore conclude that our proposed morphometric quanti-
fications show promise for use in clinical trials. Ideally, fu-
ture treatments could focus on the early stages and
thereby attempt to prevent the occurrence of severe OA.
Therefore, the ability to separate healthy from borderline
OA is essential for disease progression quantifications to
be used in clinical trials.
Table IV
Related methods for segmentation and quantification of cartilage
Paper Method (interaction time) Evaluation
Grau et al.
8
Watershed transformation (5e10 min) Good segmentation performance on four knees
Pakin et al.
9
&
Tamez-Pena et al.
11
Region growing and classification
(10e40 min)
Sparsely, Thickness intra-scan CV 3.2%
Solloway et al.
40
2D slice-wise active shape model
(1 min per slice)
Thickness intra-scan CV 4.6%
Lynch et al.
46
2D slice-wise snake (15 min) Two knees, medial femoral volume inter-scan CV 1.4%
Stammberger et al.
7
Slice-wise spline-based (2.5 h) From Koo et al.
12
: Four subjects, Thickness inter-observer CV 6.6%
Single subject, Thickness inter-scan difference 4%
From Hohe et al.
10
: Area inter-scan CV 2.5%
McWalter et al.
41
Manual outlining (1 h) Six subjects, Thickness CV 5%
Gougoutas et al.
47
&
Naish et al.
48
Original live-wire segmentation
47
extended with thickness quantification
48
(1 h per hip)
Hip, six subjects, Volume (femoral þ acetabular) inter-scan CV 2.5%
Hohe et al.
10
The segmentation from Stammberger
et al.
7
with a new curvature
quantification
16 Subjects, Curvature inter-scan CV 16%
Raynauld et al.
49
Slice-wise active contour Thorough evaluation but using measures incomparable to the others
Eckstein et al.
42
Manual outlining from 3 T DESSwe of
medial tibial cartilage
19 Subjects, inter-scan RMS CV, Volume 3.9%, Area 2.9%,
Thickness 2.3%
816
E. B. Dam et al.: Automatic morphometric cartilage quantification
Acknowledgments
A number of colleagues were very helpful with suggestions
during this work. In particular, we wish to acknowledge
Mads Nielsen, Ole Fogh Olsen, Marco Loog and Marleen
de Bruijne from the IT University of Copenhagen as well
as La
´
szlo
´
B. Tanko
´
from the Center for Clinical and Basic
Research and Morten A. Karsdal from Nordic Bioscience
in Herlev, Denmark.
References
1. Buckwalter J, Saltzman C, Brown T. The impact of os-
teoarthritis d implications for research. Clin Orthop
2004;427:S6eS15.
2. Reginster J. The prevalence and burden of arthritis.
Rheumatology 2002;41(Suppl 1):3e6.
3. Karsdal MA, Tanko LB, Riis BJ, Sondergard BC,
Henriksen K, Altman RD. Calcitonin is involved in car-
tilage homeostasis: is calcitonin a treatment for OA?
Osteoarthritis Cartilage 2006;14(7):617e24.
4. Mills G. Regulatory opportunities & challenges of imag-
ing as a drug development tool, http://www.fda.gov/
cder/regulatory/medImaging/imagingK eyNote.ppt ; 2005.
5. Abadie E, Ethgen D, Avouac B, Bouvenot G, Branco J,
Bruyere O, et al. Recommendations for the use of new
methods to assess the efficacy of disease-modifying
drugs in the treatment of osteoarthritis. Osteoarthritis
Cartilage 2004;12(4).
6. Pessis E, Drape JL, Ravaud P, Chevrot A, Ayral MDX.
Assessment of progression in knee osteoarthritis: re-
sults of a 1 year study comparing arthroscopy and
MRI. Osteoarthritis Cartilage 2003;11(5):361e9.
7. Stammberger T, Eckstein F, Michaelis M,
Englmeier KH, Reiser M. Interobserver reproducibility
of quantitative cartilage measurements: comparison
of B-spline snakes and manual segmentation. Magn
Reson Imaging 1999;17(7):1033e42.
8. Grau V, Mewes AUJ, Alcaniz M, Kikinis R, Warfield SK.
Improved watershed transform for medical image seg-
mentation using prior information. IEEE Trans Med Im-
aging 2004;23(4):447e58.
9. Pakin SK, Tamez-Pena JG, Totterman S, Parker KJ.
Segmentation, surface extraction and thickness com-
putation of articular cartilage. In: SPIE, Medical Imag-
ing, vol. 4684, 2002;155e66.
10. Hohe J, Ateshian G, Reiser M, Englmeier K, Eckstein F.
Surface size, curvature analysis, and assessment of
knee joint incongruity with MRI in vivo. Magn Reson
Med 2002;47(3):554e61.
11. Tamez-Pena JG, Barbu-McInnis M, Totterman S. Knee
cartilage extraction and boneecartilage interface anal-
ysis from 3D MRI data sets. In: SPIE, Medical
Imaging, vol. 5370, 2004;1774e84.
12. Koo S, Gold GE, Andriacchi TP. Considerations in mea-
suring cartilage thickness using MRI: factors influencing
reproducibility and accuracy. Osteoarthritis Cartilage
2005;13(9):782e9.
13. Gandy S, Dieppe P, Keen M, Maciewicz R, Watt I,
Waterton J. No loss of cartilage volume over three
years in patients with knee osteoarthritis as assessed
by magnetic resonance imaging. Osteoarthritis Carti-
lage 2002;10(12):929e37.
14. Dunn T, Lu Y, Jin H, Ries M, Majumdar S. T2 relaxation
time of cartilage at MR imaging: comparison with se-
verity of knee osteoarthritis. Radiology 2004;232(2):
592e8.
15. Kamibayashi L, Wyss U, Cooke T, Zee B. Changes in
mean trabecular orientation in the medial condyle of
the proximal tibia in osteoarthritis. Calcif Tissue Int
1995;57(1):69e73.
16. Cicuttini FM, Wluka AE, Stuckey SL. Tibial and femoral
cartilage changes in knee osteoarthritis. Ann Rheum
Dis 2001;60(10):977e80.
17. Eckstein F, Ateshian G, Burgkart R, Burstein D,
Cicuttini F, Dardzinski B, et al. Proposal for a nomen-
clature for MRI based measures of articular cartilage
in OA. Osteoarthritis Cartilage 2006;(10):14.
18. Peterfy C, Li J, Zaim S, Duryea J, Lynch J, Miaux Y,
et al. Comparison of fixed-flexion positioning with fluo-
roscopic semi-flexed positioning for quantifying radio-
graphic joint-space width in the knee: testeretest
reproducibility. Skeletal Radiol 2003;32(3):128e32.
19. Duddy J, Kirwan JR, Szebenyi B, Clarke S, Granell R,
Volkov S. A comparison of the semiflexed (MTP)
view with the standing extended view (SEV) in the ra-
diographic assessment of knee osteoarthritis in a busy
routine X-ray department. Rheumatology 2005;44(3):
349e51.
20. Xia Y. The total volume and the complete thickness of
articular cartilage determined by MRI. Osteoarthritis
Cartilage 2003;11(7):473e4.
21. Kellgren JH, Lawrence JS. Radiological assessment of
osteo-arthrosis. Ann Rheum Dis 1957;16(4):494e501.
22. Folkesson J, Dam EB, Olsen OF, Pettersen PC,
Christiansen C. Segmenting articular cartilage auto-
matically using a voxel classification approach. IEEE
Trans Med Imaging 2007;26:106e 15.
23. Folkesson J, Dam EB, Olsen OF, Pettersen PC, Christi-
ansen C. Position normalization in automatic cartilage
segmentation. In: MICCAI Joint Disease Workshop,
2006;9e16. Available at http://www.diku.dk/werikdam/
JOINT.
24. Dam EB, Folkesson J, Loog M, Pettersen PC, Christi-
ansen C. Efficient automatic cartilage segmentation.
In: MICCAI Joint Disease Workshop, 2006;88e95.
Available at http://www.diku.dk/werikdam/JOINT.
25. Joshi S, Pizer S, Fletcher PT, Yushkevich P, Thall A,
Marron JS. Multiscale deformable model segmenta-
tion and statistical shape analysis using medial de-
scriptions. Trans Med Imaging 2002;21(5):538e50.
26. Dam EB, Folkesson J, Pettersen PC, Christiansen C.
Automatic cartilage thickness quantification using
a statistical shape model. In: MICCAI Joint Disease
Workshop, 2006;42e9. Available at http://www.diku.
dk/werikdam/JOINT.
27. Williams T, Holmes A, Maciewicz R, Waterton J,
Taylor C, Creamer P, et al. Cartilage loss in osteoar-
thritis detected by statistical shape analysis of mag-
netic resonance images. Osteoarthritis Cartilage
2005;13(Suppl A).
28. Folkesson J, Dam EB, Olsen OF, Pettersen PC, Christi-
ansen C. Automatic curvature analysis of the articular
cartilage surface. In: MICCAI Joint Disease Workshop,
2006;17e24. Available at http://www.diku.dk/w
erikdam/JOINT.
29. Vilalta C, Nunez M, Segur JM, Carbonell ADJA,
Macule F. Knee osteoarthritis: interpretation variability
of radiological signs. Clin Rheumatol 2004;23(6):
501e4.
30. Peterfy CG, Guermazi A, Zaim S, Tirman PFJ, Miaux Y,
White D, et al. Whole-Organ Magnetic Resonance
Imaging Score (WORMS) of the knee in osteoarthritis.
Osteoarthritis Cartilage 2004;12(3).
817
Osteoarthritis and Cartilage Vol. 15, No. 7
31. Millington S, Li K, Wu X, Hurwitz S, Sonka M. Auto-
mated simultaneous 3D segmentation of multiple car-
tilage surfaces using optimal graph searching on MRI
images (Abstract). Osteoarthritis Cartilage 2005;
13(Suppl A):S130.
32. Trinh NH, Lester J, Fleming BC, Tung G, Kimia BB. Ac-
curate measurement of cartilage morphology using
a 3D laser scanner. In: Proceedings of CVAMIA,
Workshop at European Conference on Computer
Vision. Springer 2006;37e48.
33. Graichen H, Jakob J, von Eisenhart-Rothe R,
Englmeier KH, Reiser M, Eckstein F. Validation of
cartilage volume and thickness measurements in
the human shoulder with quantitative magnetic reso-
nance imaging. Osteoarthritis Cartilage 2003;11(7):
475e82.
34. Kauffmann C, Gravel P, Godbout B, Gravel A,
Beaudoin G, Raynauld J, et al. Computer-aided
method for quantification of cartilage thickness and
volume changes from RMI: validation study using
a synthetic model. IEEE Trans Biomed Eng 2003;50(8).
35. Eckstein F, Cicuttini F, Raynauld JP, Waterton JC,
Peterfy C. Magnetic resonance imaging (MRI) of artic-
ular cartilage in knee osteoarthritis (OA): morphologi-
cal assessment. Osteoarthritis Cartilage 2006;
14(Suppl 1):46e75.
36. Ding C, Garnero P, Cicuttini F, Scott F, Cooley H,
Jones G. Knee cartilage defects: association with early
radiographic osteoarthritis, decreased cartilage volume,
increased joint surface area and type II collagen break-
down. Osteoarthritis Cartilage 2005;13(3):198e205.
37. Jones G, Ding C, Scott F, Glisson M, Cicuttini F. Early
radiographic osteoarthritis is associated with substan-
tial changes in cartilage volume and tibial bone sur-
face area in both males and females. Osteoarthritis
Cartilage 2004;12(2):169e74.
38. Amin S, LaValley MP, Guermazi A, Grigoryan M,
Hunter DJ, Clancy M, et al. The relationship between
cartilage loss on magnetic resonance imaging and ra-
diographic progression in men and women with knee
osteoarthritis. Arthritis Rheum 2005;(10):52.
39. Terukina M, Fujioka H, Yoshiya S, Kurosaka M,
Makino T, Matsui N, et al. Analysis of the thickness
and curvature of articular cartilage of the femoral con-
dyle. Arthroscopy 2003;19.
40. Solloway S, Hutchinson C, Waterton J, Taylor C. Quan-
tification of articular cartilage from MR images using
active shape models. In: European Conference on
Computer Vision. Springer 1996;400e11.
41. McWalter EJ, Wirthz W, Siebert M, von Eisenhart-
Rothe RMO, Hudelmaier M, Wilson DR, et al. Use of
novel interactive input devices for segmentation of ar-
ticular cartilage from magnetic resonance images. Os-
teoarthritis Cartilage 2005;13:48e53.
42. Eckstein F, Hudelmaier M, Wirth W, Kiefer B,
Jackson R, Yu J, et al. Double echo steady state mag-
netic resonance imaging of knee articular cartilage at 3
Tesla: a pilot study for the osteoarthritis initiative. Ann
Rheum Dis 2006;65(4):433e41.
43. Peterfy CG, Gold G, Eckstein F, Cicuttini F,
Dardzinski B, Stevens R. MRI protocols for whole-or-
gan assessment of the knee in osteoarthritis. Osteoar-
thritis Cartilage 2006;14(Suppl 1):95e111.
44. Huegli RW, Tirman PFJ, Bonel HM, Staedele H,
Zaim S, Grigorian M, et al. Use of the modified
three-point Dixon technique in obtaining T1-weighted
contrast-enhanced fat-saturated images on an open
magnet. Eur Radiol 2004;14:1781e6.
45. Dam EB, Folkesson J, Pettersen PC, Christiansen C.
Semi-automatic knee cartilage segmentation. In:
Proc. SPIE Medical, 2006.
46. Lynch JA, Zaim S., Zhao J, Stork A, Peterfy CG,
Genant HK. Automatic measurement of subtle
changes in articular cartilage from MRI of the knee
by combining 3D image registration and segmentation.
In: SPIE, Medical Imaging, vol. 4322, 2001;431e9.
47. Gougoutas A, Wheaton A, Borthakur A, Shapiro E,
Kneeland J, Udupa J. Cartilage volume quantification
via live wire segmentation. Acad Radiol 2004;11(12):
1389e95.
48. Naish JH, Xanthopoulos E, Hutchinson CE, Waterton JC,
Taylor CJ. MR measurement of articular cartilage
thickness distribution in the hip. Osteoarthritis Carti-
lage 2006;14(10):967e73.
49. Raynauld JP, Kauffmann C, Bea udoin G, Berthiaume MJ,
de Guisei JA, Bloch DA,
et al. Reliability of a quantifica-
tion imaging system using magnetic resonance images
to measure cartilage thickness and volume in human
normal and osteoarthritic knees. Osteoarthritis Carti-
lage 2003;11:351e60.
818
E. B. Dam et al.: Automatic morphometric cartilage quantification
    • "For the experiments described in this thesis, we used data sets consisting of MRI scans of both left and right knees from 159 test subjects in a longitudinal, community-based, non-treatment study [33]. After exclusion of scans due to acquisition artefacts, 313 knee scans remained in the diagnosis data set. "
    [Show abstract] [Hide abstract] ABSTRACT: The pathogenesis of osteoarthritis (OA) includes complex events in the whole joint. Cartilage loss and bone remodelling are central in OA progression. In this project, we investigated the feasibility of quantifying OA by analysis of the tibial trabecular bone structure in low-field knee magnetic resonance imaging (MRI). The development of automatic and more sensitive indicators of OA in conjunction with low cost equipment have the potential to decrease the length and cost of clinical trials. We present a texture analysis methodology that combined uncommitted machine-learning techniques in a fully automatic framework. Different linear feature selection approaches where investigated. The methodology was evaluated in a longitudinal study, where MRI scans of knees were used to quantify the tibial trabecular bone in a bone marker for OA diagnosis and another marker for prediction of tibial cartilage loss. The healthy and diseased subjects were defined by the Kellgren and Lawrence index assigned by radiologists and the levels of cartilage loss were assessed by a segmentation process. A preliminary radiological reading of the knees with high and low risks of cartilage loss suggested the prognosis bone marker captured aspects of the vertical trabecularization of the tibial bone to define the prognosis of cartilage loss. We also investigated which region of the tibia provided the best prognosis for medial tibial cartilage loss. The structure of the tibial trabecular bone was divided in localized subregions in an attempt to capture the different pathological features occurring at each location. We applied multiple-instance learning, where each subregion was defined to be one instance and a bag held all instances over a full region-of-interest. The inferior part of the tibial bone was classified as the most relevant region for prognosis of cartilage loss and a preliminary radiological reading of a subset of the samples suggested the bone marker also captured the vertical trabecularization of the tibial bone to define the most relevant region. In a clinical point of view, besides presenting a bone marker able to predict disease progression and diagnostic bone marker superior to other OA biomarkers, our findings underlined the importance of the trabecular bone to the understanding of the OA pathology.
    Full-text · Thesis · Apr 2013
    • "However, several research groups are still trying to determine which cartilage morphometric measure is most responsive or valid in predicting progression (Eckstein et al., 2006;). Several studies have also established the superiority of using MR at 3.0T for measuring cartilage thickness and volume in an accurate way (Kshirsagar et al., 1998; Raynauld, 2003; Eckstein et al., 2006; Dam et al., 2007; Guermazi et al., 2008). Eckstein et al. performed a study to measure thickness, volume, and surface area of the femorotibial cartilage. "
    Full-text · Chapter · Feb 2012
    • "The use of peripheral MRI scanners at lower field strengths potentially permits more widespread distribution of this technology, especially when access to high-field MRI is limited. Quantitative cartilage measurement at 0.2 T have also been proposed [18, 27, 28,383940 but have not been validated versus external standards or measurement at higher field strength. However, they were shown to display substantially larger precision errors than measurements performed at higher field strength. "
    [Show abstract] [Hide abstract] ABSTRACT: Quantitative measures of cartilage morphology (i.e., thickness) represent potentially powerful surrogate endpoints in osteoarthritis (OA). These can be used to identify risk factors of structural disease progression and can facilitate the clinical efficacy testing of structure modifying drugs in OA. This paper focuses on quantitative imaging of articular cartilage morphology in the knee, and will specifically deal with different cartilage morphology outcome variables and regions of interest, the relative performance and relationship between cartilage morphology measures, reference values for MRI-based knee cartilage morphometry, imaging protocols for measurement of cartilage morphology (including those used in the Osteoarthritis Initiative), sensitivity to change observed in knee OA, spatial patterns of cartilage loss as derived by subregional analysis, comparison of MRI changes with radiographic changes, risk factors of MRI-based cartilage loss in knee OA, the correlation of MRI-based cartilage loss with clinical outcomes, treatment response in knee OA, and future directions of the field.
    Full-text · Article · Jan 2011
Show more

  • undefined · undefined
  • undefined · undefined
  • undefined · undefined