106IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 1, JANUARY 2007
Segmenting Articular Cartilage Automatically Using
a Voxel Classification Approach
Jenny Folkesson, Member, IEEE, Erik B. Dam, Ole F. Olsen, Member, IEEE, Paola C. Pettersen, and
Abstract—We present a fully automatic method for articular
cartilage segmentation from magnetic resonance imaging (MRI)
which we use as the foundation of a quantitative cartilage as-
sessment. We evaluate our method by comparisons to manual
segmentations by a radiologist and by examining the interscan
reproducibility of the volume and area estimates. Training and
evaluation of the method is performed on a data set consisting of
139 scans of knees with a status ranging from healthy to severely
osteoarthritic. This is, to our knowledge, the only fully automatic
cartilage segmentation method that has good agreement with
manual segmentations, an interscan reproducibility as good as
that of a human expert, and enables the separation between
healthy and osteoarthritic populations. While high-field scanners
offer high-quality imaging from which the articular cartilage have
been evaluated extensively using manual and automated image
analysis techniques, low-field scanners on the other hand produce
lower quality images but to a fraction of the cost of their high-field
counterpart. For low-field MRI, there is no well-established accu-
racy validation for quantitative cartilage estimates, but we show
that differences between healthy and osteoarthritic populations
are statistically significant using our cartilage volume and surface
area estimates, which suggests that low-field MRI analysis can
become a useful, affordable tool in clinical studies.
Index Terms—Articular cartilage, image segmentation, os-
teoarthritis, magnetic resonance imaging (MRI), pattern classifi-
in causing work disability and is associated with a large socioe-
conomic impact on health care systems . One of the main ef-
fects of OA is the degradation of the articular cartilage, causing
pain and loss of mobility of the joints. Currently, the treatment
of OA is mainly restricted to symptom control , and in the
search for disease modifying drugs, much research is dedicated
to analysis of articular cartilage and its relation to disease pro-
Magnetic resonance imaging (MRI) is the leading imaging
modality for direct, noninvasive assessment of the articular
cartilage , and cartilage deterioration can be detected using
quantitative MRI analysis . Among MRI sequences, the
STEOARTHRITIS (OA) is one of the major health issues
among the elderly population, it is second to heart disease
Manuscript received June 27, 2006; revised September 15, 2006. Asterisk
indicates corresponding author.
S, Denmark (e-mail: firstname.lastname@example.org).
E. B. Dam and O. F. Olsen are with the IT University of Copenhagen,
DK-2300 Copenhagen S, Denmark.
P. C. Pettersen and C. Christiansen are with Center for Clinical and Basic
Research, 2750 Ballerup, Denmark.
Digital Object Identifier 10.1109/TMI.2006.886808
most established are fat-suppressed gradient-echo T1 se-
quences using a 1.5T or a 3T magnet. The standard sequences
in literature for these scanners have high in-plane resolution but
usually have a larger interslice distance, and many assessment
methods developed for such sequences are on a slice-by-slice
basis. For a thorough review of MRI scan protocols for knee
OA assessment, see .
A recent study shows that low-field dedicated extremity MRI
can provide similar information on bone erosions and synovitis
as expensive high-field MRI units . There have been several
comparisons of diagnostic performance of diagnosing meniscal
andhigh-field MRIdata –reportingeverythingfrom com-
patible performance to the high-field unit outperforming the
low-field unit. There has also been a comparison between low-
field MRI and arthroscopy  finding a good correspondence
between the two for cruciate ligament and lesion detection in
The use of a dedicated low-field MRI has its advantages
and disadvantages. The drawbacks are related to image quality
with lower resolution and more difficulties in incorporating
features such as fat suppression, however fat suppression has
been successfully implemented lately for low-field MRI .
The main advantages are cost-effectiveness with much lower
cost per scan, lower installation and maintenance costs, and
higher patient comfort without claustrophobic feelings and
minimal noise level. So far there has not been any validation
of quantitative cartilage measures from a low-field scanner
compared to ground truth, but if a low-field scanner can be used
for quantitative articular cartilage assessment, costs for making
clinical studies would be reduced significantly. If manual labor
is connected with the analysis and quantification of MRI data
in clinical studies, one more cost factor is introduced. In this
assessment framework, and we evaluate it on low-field MRI by
comparison to manual delineations by a radiologist, we eval-
uate the robustness in terms of interscan reproducibility, and
the ability to detect changes between healthy and osteoarthritic
groups using the cartilage volume and area estimates.
A. Related Work
As in most quantitative assessment studies in medical
imaging, the first and most crucial step in our articular cartilage
assessment is segmentation. The cartilage can be manually
segmented slice-by-slice by experts, but for routine clinical use
manual methods are too time consuming and they are prone
to inter- and intraobserver variability. It is thus advantageous
to automate the segmentation method and the main challenges
0278-0062/$20.00 © 2006 IEEE
FOLKESSON et al.: SEGMENTING ARTICULAR CARTILAGE AUTOMATICALLY107
in developing an automatic method are the thin structure of
the cartilage and the low contrast between the cartilage and
surrounding soft tissues.
Several groups have developed semiautomated/automated
methods for cartilage segmentation. Among two-dimensional
(2-D) techniques, Stammberger and colleagues  segments
the cartilage by fitting a -spline snake to each slice. A 2-D
method combining user interaction with active contours is
described by Lynch et al. . They combine the segmentation
technique with three-dimensional (3-D) image registration to
detect changesin cartilagevolume.Sollowayetal. use
active shape models for slice-by-slice cartilage segmentation,
and estimate cartilage thickness in the direction perpendicular
to the medial axis in each slice.
When working with a 2-D technique, continuation between
slices is lost and some regularization between the slices is re-
quired. Also, since the series of 2-D segmentations have to be
converted into a 3-D segmentation when finding for example
thickness maps, it is advantageous to perform segmentation in
Looking at the 3-D techniques that have been developed,
Grau et al.  use a watershed based approach, where the
watershed is extended to examining difference in class prob-
ability of neighboring pixels. The sensitivity, specificity and
Dice volume overlap of the segmentation are 90.03%, 99.86%,
and 0.90, respectively. The method is evaluated on seven scans
from four healthy knees and requires 5–10 min of manual labor
for selecting markers before initializing the watershed.
Pakin et al.  has developed a region growing scheme that
is followed by a two-class clustering for segmenting the carti-
lage. However, the method assumes that the bones are already
segmented. The sensitivity and specificity of the method are
66.22% and 99.56%, respectively, and it is evaluated on one
scan. The method has been further developed to incorporate
a trained user for correcting misclassifications , and this
semiautomatic method is evaluated in terms of intrauser repro-
Another classification approach to segmentation is presented
by Warfield et al. , , where a user performs interactive
registration of a knee template to a test scan. The method then
iterates between a classification step and a template registration
step to produce a segmentation. The method has a lower in-
trascan variability of the volume compared to repeated manual
segmentations on the scan it is evaluated on.
tation algorithm  followed by mean thickness quantification
is evaluated on ankle joints in . The method requires only a
small amountof manual initialization and shows accuratethick-
ness measurements on eight cadaveric ankles. Presumably, the
method could also be adapted to knees.
B. Overview of the Work Presented
The segmentation techniques described in Section I-A all re-
quire some amount of manual interaction except for the method
of Pakin et al. , the 3-D techniques are evaluated only on
relatively small data sets and neither Grau et al. nor Pakin et al.
evaluate their methods on scans from OA test subjects.
In this paper, we propose a method that can fully automati-
cally segment cartilage in both healthy and osteoarthritic knee
fully automatic cartilage assessment and is primarily intended
for clinical studies using low-field MR scanners. The segmen-
tation algorithm is based on a one versus all approach of com-
bining binary approximate NN classifiers which is described
in Sections III-A and III-B, followed by an iterative position
adjustment method that is intended to correct for the variations
of the placement of the test subject in the scanner, something
that is bound to occur in any clinical study and is described in
Section III-G. Since
NN classification is a slow process we
propose to use an efficient voxel classification algorithm which
is described in Section III-F.
Since we cannot obtain ground truth for an in vivo study
with both healthy and OA test subjects and ground truth ac-
curacy of low-field MRI analysis is yet to be established, we
evaluate our method not only compared to manual tracings of a
radiologist, but also in terms of precision. We evaluate the in-
terscan reproducibility using the volume and surface area esti-
mate, and the ability to detect changes between healthy and os-
teoarthritic populations by performing unpaired -tests between
the groups using the volume and area estimates and the Kell-
gren–Lawrence index. OA is more frequently observed in the
tilage compartment in this study. The evaluation of the segmen-
tation framework is described in Section IV followed by discus-
sion in Section V.
II. IMAGE ACQUISITION
A. Magnetic Resonance Image Acquisition
MRI was performed with an Esaote C-Span lowfield 0.18T
scanner dedicated to imaging of extremities yielding a sagittal
Turbo 3-D T1 sequence (40 flip angle,
Approximate acquisition time is 10 min and the scan size, after
voxels. The spatial in-plane resolution of
the scans are 0.70
0.70 mm , with a distance between slices
ranging between 0.70 mm–0.94 mm, where the most common
distance is 0.78 mm.
of limited continuation between slices that is present in 2-D
techniques. We use a 3-D sequence consisting of near isotropic
voxels since this is well suited for cartilage quantification 
and for 3-D analysis in general.
50 ms,16 ms).
B. Test Subject Population
We examine 139 knee joints in vivo, of which 59% are from
female test subjects. The ages of the test subjects varies be-
tween 22–79 years with an average age of 56 years. The status
of the knees range from healthy to osteoarthritic according to
the Kellgren–Lawrence index
established by X-rays between 0–4 where
is considered borderline or mild OA, and
is severe OA. In our data set, 51 knees have
haveand the remaining 22 knees have
. In the X-rays the width of the tibial plateau has also
, a radiographic score
, 28 have
108IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 1, JANUARY 2007
Fig. 1. Scan most improved by the position correction scheme, where the DSC increases from 0.61 to 0.77. First column shows the manual segmentation, the
second column shows the original segmentation, and the third column shows the segmentation after position correction. 2-D images in the top row are a sagittal
slice of the segmentation and the 3-D views on the second row are of the same segmentation seen from above.
Fig. 2. Worst case scenario of applying position correction. Knee is severely osteoarthritic ???? ? ??. For this scan, there is no improvement in DSC. Manual
segmentation is in the first column, the second column shows initial segmentation, and the third column shows the segmentation after position correction. 2-D
images in the top row are a sagittal slice of the segmentation and the 3-D views on the second row are of the same segmentation seen from above.
been measured, which we use for normalization of the cartilage
sizes can be compared. The scans are from both left and right
knees, and in order to treat all scans analogously with the same
methods, all the right knees are reflected about the center of the
The images are transmitted from the MRI unit to a worksta-
tion, where they are processed using a medical imaging display
and analysis system designed for the task. The software allows
for manual segmentation on a slice-by-slice basis. A user marks
points on the object boundary, and linear interpolation between
the points delineates the boundary. The MR scans have all been
manually segmented by a radiologist using this software, and
31 scans are segmented twice with the purpose of examining
the intrarater variability of the manual delineations.
Of the 139 knees, the same 31 knees that were segmented
twice were rescanned after approximately one week in order to
examine the segmentation precision, giving a total of 164 MR
scans. An example of how a MRI slice and the manual delin-
eation looks like can be seen in the first column of Figs. 1 and 2.
III. CARTILAGE SEGMENTATION
A. Voxel Classification
We implement our classifier in an approximate nearest
neighbor framework developed by Mount and colleagues .
The classifier is in principle a
faster computationsifan error inthedistancecalculations istol-
erated. The approximate
search algorithm returns
NN-classifier, but allows for
FOLKESSON et al.: SEGMENTING ARTICULAR CARTILAGE AUTOMATICALLY109
such that the ratio of the distance between the th reported point
and the true th nearest neighbor is at most
Given a set
ofdata points in
of a point in
can be computed in
, thus the computational complexity
increases exponentially with the dimensions. One difficulty
in classification tasks is the tradeoff between computational
complexity and accuracy. We found empirically that
give a reasonable such tradeoff.
, thenearest neighbors
B. Multiclass Classification by Combining Binary Classifiers
There are three classes we wish to separate, tibial medial car-
tilage, femoral medial cartilage and background. We combine
one binary classifier trained to separate tibial cartilage from
the rest and one trained to separate femoral cartilage from the
rest with a rejection threshold
a one vs. rest classifier can be seen as the posterior probabil-
ities that, for all the voxels in the image, a voxel
belongs to class
number of classes. We denote it
In one-versus-all classification, which is commonly used for
multi-class classification , one builds
sifiers and perform a winner-takes-all vote between them, as-
to the classwith the highest posterior probability.
and 0.5% to the femoral cartilage, making the background the
by far largest class. Our approach is similar to one-versus all,
but due to the dominance of the background class we replace
the background versus rest classifier by a rejection threshold,
which states that the posterior probability should be higher than
the threshold before it can be assigned to a cartilage class. The
decision rule is
, . The outcome of
one vs. rest clas-
tibial medial, femoral medial and background, respectively.
The rejection threshold is optimized on the training set to max-
imize the dice similarity coefficient (DSC) which is considered
a useful statistical measure for studying agreement between
different segmentations . It measures the spatial volume
overlap between two segmentations
and the subscriptsandstands for
andand is defined as
C. Feature Selection
Feature selection can provide a suitable feature set for the
classification task at hand. The features of the classifiers are
selected by sequential forward selection followed by sequen-
tial backward selection from a large bank of features described
below in Section III-D, . In the forward selection, we start
one feature at the time according to the outcome of a crite-
(ROC) curve . The backward selection starts with the fea-
tures found by the sequential forward selection and iteratively
excludes the least significant feature according to the criterion
All features are examined in every iteration which means that
tablish an indirect weighting of important features. We use 25
in the feature selection, threshold selection and for the training
data set for the final classifier. Using 25 scans gives us a large
enough training set to not be sensitive to the curse of dimen-
sionality—the outcome of the criterion function evaluation is
improved after every iteration and we stop iterating when the
effective due to the exponential increase in computational com-
plexity with the number of dimensions. We do backward selec-
tion until there are 39 features remaining in the set, and we ob-
served that for these iterations there is no significant decrease in
the classifier performance. This feature selection scheme does
not guarantee a global optimum, but by doing forward selection
followed by backward selection we search a larger part of the
tree consisting of all possible combinations of features (given
or less determined by computational complexity) than by only
using forward selection.
Wecombinebinaryclassifierseventhough NN is inherently
a multiclass classifier. The reason for so doing is that for feature
selection, the area under the ROC curve evaluates the classifier
performance for all operating points for a two-class task. But
there is no obvious extension of ROC analysis for multiclass
classification tasks and we have found better results by training
and combining binary classifier than we have with direct multi-
class classifiers .
We here introduce the set of candidate features from which
the feature selection scheme selects a subset.
The intensity and the position in the image are both fea-
tures that are highly relevant for a radiologist when visually
inspecting a scan, and that is the main motivation for including
them as candidate features. Both the raw image intensities
and intensities from the image convolved with a Gaussian
according to the scale space framework  on different scales
are considered. Three scales are chosen (0.65, 1.1, and 2.5 mm)
to cover the range of different cartilage thicknesses. Though
the location and the shape of the cartilage varies from scan to
scan, the coordinates are still an indicator of where cartilage is
more likely to be situated.
Other features of interest are those related to the geometry
of the object in question. The three-jet, which consists of all
forms a basis which describes all geometric features up to third-
order  and are thus considered as candidate features. The
-, -, and -axes are here defined as the sagittal-, coronal-, and
It is well known that numerical differentiation enhances
higher spatial frequencies and that the effect increases with the
order of the differentiation, meaning that noise may limit the
practical use of higher order derivatives. Blom  shows that
110 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 1, JANUARY 2007
the spatial averaging in scale-space causes a noise reduction
that more than counteracts the noise amplification caused by
differentiation. Hence, all the derivatives mentioned in this
section are achieved by convolution with Gaussian derivatives,
a differential operator, and
derivatives or smoothing are involved are examined at the three
different scales mentioned above. Though the lowest scale we
use (0.65 mm) is lower than the resolutions of the scans, there
is still some spatial averaging and Gaussian derivatives allow
for robust differentiation at that scale.
In vessel segmentation, the eigenvalues of the Hessian (H)
have proven to be useful when looking for central locations in-
side a tubelike structure . The Hessian is the symmetric ma-
trix containing second-order derivatives with respect to the co-
, whereis a Gaussian,
is the scale. All features where
and it describes the second-order structure of local inten-
sity variations. The largest eigenvalue gives the highest
second-order derivative value and its corresponding eigen-
vector is in the direction of the maximum second-order
derivative. Cartilage can locally be described as a thin disc,
which corresponds to finding positions with one large and two
small eigenvalues of the Hessian. The eigenvalues and the three
eigenvectors are candidate features.
tion of thin structures such as fingerprints is the structure tensor
( ) . The structure tensor is a symmetric matrix containing
where the outer scale
the one used for obtaining the derivatives
tensor examines the local gradient distribution at each location
. The directions of the eigenvectors depend on the vari-
ation in the neighborhood. The structure tensor eigenvalues and
eigenvectors combining three different scales on
ture for the local third-order structure as well. The third-order
derivatives with respect to
sented in the third-order tensor
structure in the local gradient direction
scribed using Einstein summation as
is not necessarily the same scale as
. The structure
can be conveniently repre-
. Examining the third-order
can be de-
The third-order tensor examined in the gradient direction on
three different scales are candidate features.
In summary, our candidate features are the intensity, the po-
sition, the three-jet, eigenvalues, and eigenvectors of both the
Hessian and the structure tensor and the third-order tensor in
the gradient direction. All features except the position are cal-
culated at three different scales (0.65, 1.1, and 2.5 mm), and the
scales are in mm instead of number of voxels for handling scans
with different resolutions.
All features except the intensity are coupled three by three to
allow them the same odds of getting picked. The three by three
grouping comes natural because we have3-D images, so the co-
ordinates, first-order derivatives and the eigenvalues and eigen-
vectors of the Hessian and the structure tensor have a natural
grouping. The other features are grouped using the three scales.
E. Selected Features
image, the intensities smoothed on the three scales,
three scales, the first-order derivatives on the three scales,
on the three scales, the eigenvalues of
three scales, the eigenvalues of
(2.5 mm, 0.65 mm).
versus. rest classifier contains the following fea-
tures after feature selection: the position, the eigenvector cor-
responding to the largest eigenvalue of
the first-order derivatives on scales 1.1 mm and 2.5 mm, the in-
tensity smoothed on three scales,
on all three scales, the eigenvalues of the Hessian on all three
scales, and the eigenvalues of
It can be noted that the position is selected as the most sig-
nificant feature by both classifiers. The intensity smoothed on
three scales is also ranked high by both classifiers, followed by
eigenvalues of both the Hessian and the structure tensor on var-
ious scales and second- and third-order derivatives in the direc-
tion of the coronal and axial axes.
(1.1 mm),on all
(2.5 mm), and the eigenvalues
(1.1 mm, 0.65 mm),
on the three scales,
(2.5 mm, 0.65 mm).
F. Efficient Voxel Classification
Our segmentation method is fully automatic, but due to the
high computational complexity of the
takes approximately 60 min to classify all voxels in a scan
consisting of around two million voxels by the two binary
classifiers. Even though computation power is relatively in-
expensive, such long computation times are inconvenient in
clinical studies using large numbers of scans.
algorithm , and the basic idea behind it is to not classify all
voxels but to focus mainly on the cartilage voxels. The algo-
rithm is conceptually very simple: starting from a set of ran-
domly sampled voxels, we classify them as either cartilage or
background. If a voxel is classified as cartilage, we continue
with classification of the neighboring voxels and this expansion
process continues until no more cartilage voxels are found.
This results in a number of connected regions of cartilage.
Provided that our initial sampling of starting voxels hits each
cartilage sheet in at least a single voxel, the resulting segmen-
tation will be exactly like the one resulting from a full voxel
classification after extraction of the largest connected compo-
nent. This is ensured by not making the initial random sampling
NN classification it
FOLKESSON et al.: SEGMENTING ARTICULAR CARTILAGE AUTOMATICALLY111
RESULTS FROM OUR AUTOMATIC SEGMENTATION METHOD BEFORE AND AFTER POSITION ADJUSTMENT (PA) FOR MEDIAL TIBIAL, MEDIAL FEMORAL, AND THE
MEDIAL COMPARTMENTS TOGETHER. SENSITIVITY, SPECIFICITY, AND DSC ARE FOUND FROM COMPARISON WITH MANUAL SEGMENTATIONS ON THE 114 SCANS
IN THE TEST SET. STANDARD DEVIATIONS ARE DENOTED SD AND 95% CONFIDENCE INTERVALS ARE DENOTED CI
too sparse. Since some parts of the cartilage compartments will
be fairly centered in scan we sample fairly densely at the center,
with a sampling probability of 5% for each voxel, and gradually
more sparsely towards the periphery.
G. Position Adjustment
Besides a large anatomical variation, the placement of the
knee in the scanner in clinical studies is a source of variation.
tilage, which is evident in our segmentation method where the
position is selected as one of the most significant features. Even
though the global location is a strong cue the minor variation in
placement is a source of errors. Segmentation methods that rely
on manual interaction are usually less sensitive to knee place-
ment since a user can define where in the scan the cartilage is.
We, however, have a segmentation technique that is completely
independent of user interaction thus the placement variations
that occur in scans in clinical studies is an issue that needs at-
One way of correcting for knee placement is to manually de-
termine where in the scan the cartilage is, but this can take time
with 3-D images since a human expert typically search through
the scans on a slice-by-slice basis. And when the segmentation
method itself is automatic, an automatic adjustment is advanta-
In order to adjust the segmentation method to become more
robust to variations in knee placement we have developed an
iterative scheme which consists of two steps that are repeated
until convergence . The first step consists of shifting the co-
ordinates of the scan so that the cartilage center of mass found
of mass for the cartilage points in the training set. Then in the
second step the scan is classified using the sample expand algo-
rithm with the other features unchanged. The outcome is com-
bined according to (1) and the largest connected component is
selected as the cartilage segmentation.
The position of the tibial and femoral compartments are
shifted individually for the two binary classifiers because the
classification depends on the training set, and there the different
cartilage compartments have different relative position with
respect to each other due to different positions of the test
The average computation time for automatic segmentation
of a scan is approximately 10 minutes on a standard desktop
2.8-Ghz PC. For a trained radiologist it takes around two hours
to segment the tibial and femoral medial cartilage in a scan with
slice by slice delineation of the contour by manual selection of
boundary points and automatic linear interpolation.
A. Comparison Between Automatic andManual Segmentations
The methods are trained on 25 scans and evaluated on 114
scans. Of the 114, 31 knees have been rescanned and the repro-
ducibility is evaluated by comparing the volume and area esti-
mates from the first and second scanning.
Before applying the position adjustment scheme described
in Section III-G, the automatic segmentation method yields an
average sensitivity, specificity and DSC of 81.1%, 99.9%, and
0.79, respectively, for the total medial cartilage segmentation,
in comparison with manual segmentations.
After applying the automatic position normalization, the av-
erage sensitivity, specificity, and DSC are 83.9%, 99.9%, and
0.80, respectively. The scheme converges in only one iteration.
Compared to the initial segmentation there is a significant in-
crease in sensitivity
according to a paired -test. In order to illustrate how the
segmentations are affected, the best and worst cases from the
position correction scheme are shown in Figs. 1 and 2. In the
best case, the DSC increases with 0.17 and for the worst scan
it decreases with 0.017. The scan with the worst result is from
a severely osteoarthritic knee which can be difficult even for a
highly trained expert to segment. The results for each compart-
ment is listed in Table I.
When comparing between manual and automatic estimates
for the 114 scans, the average pairwise differences for medial
volume and area are 8.7% and 0.05%, respectively. The volume
from the automatic method overestimates the manual with 10%
with significant difference between group means
and the area is underestimated by 0.7% with no significant
. Some of the overestimation of the
volume most likely originates from false positives from lateral
and patellar cartilage that is adjacent to the medial compart-
ments. Visual inspection supports this, for instance in Fig. 1 it
can be seen that the manual segmentation ends more abruptly
at the medial/lateral border than the automatic segmentation.
Still this remains to be verified statistically in a future study
including all compartments. Also, there is an uncertainty in the
segmentation close to the boundary particularly along the crest,
and the scans used in this study have low contrast between
tissues which may also contribute to false positives compared
to manual segmentations.
and in DSC
112IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 1, JANUARY 2007
INTERSCAN REPRODUCIBILITY OF OUR AUTOMATIC SEGMENTATION METHOD
BEFORE AND AFTER POSITION ADJUSTMENT (PA) AND OF THE MANUAL
SEGMENTATIONS (M), FOR MEDIAL TIBIAL, MEDIAL FEMORAL, AND THE
MEDIAL COMPARTMENTS TOGETHER. LINEAR CORRELATION COEFFICIENT
(CORR.) AND AVERAGE ABSOLUTE PAIRWISE DIFFERENCES (DIFF.) FOR THE
31 KNEES SCANNED TWICE
Fig. 3. Bland–Altman plot of the interscan reproducibility of the tibial volume
from automatic (position adjusted) segmentations. Lines are the mean ?? SD
of the difference between measurements.
from the automatic segmentations, we examine the 31 knees
that were scanned twice. Before position adjustment there is an
average absolute volume and area difference of 10% and 6.0%
for the total medial cartilage, and after position adjustment the
reproducibility of the method is improved, with a decrease of
the average absolute volume and area differences to 6.5% and
4.5% respectively. These values can be compared to the repro-
ducibility of the manual segmentation which has an average ab-
solute volume and area difference of 6.5% and 5.5% respec-
tively for the same data set. The reproducibility for the auto-
matic method and human expert for both volume and area for
all compartments are listed in Table II, where it can be seen
that the tibial volume and area estimates are the most repro-
ducible for the automatic method, possibly because the tibial
cartilage has a less complex shape compared to the femoral car-
tilage. In Figs. 3 and 4, the Bland–Altman plots of interscan
reproducibility for the automatically obtained tibial volume and
area estimates are displayed.
The radiologist has a fairly poor precision on volume both
tibial and femoral separately, but it improves when the two
compartments are combined. This shows that the radiologist is
mainly in doubt on the part of the cartilage sheets where tibial
Fig. 4. Bland–Altman plot of the interscan reproducibility of the tibial area
from automatic (position adjusted) segmentations. Lines are the mean ?? SD
of the difference between measurements.
and femoral are touching. These volume precision numbers are
lower than what is reported in other studies, something which
could be a consequence of the low-field low resolution scans
used in this study.
tilage in 31 scans in order to determine intrarater variability for
the manual segmentations. The average DSC between the two
manual segmentations are 0.87 for the medial cartilage, which
explains the fairly low values of the DSC in our evaluation be-
cause the method is trained on manual segmentations by the ex-
pert and therefore attempts to mimic the expert. Also, assuming
most misclassifications occur at boundaries, thin structures will
typically have relatively low DSC. The corresponding DSC of
the automatic segmentation versus expert for the medial carti-
lage of the 31 scans is 0.80.
For all the scans the in-plane resolution is 0.70
but the slice distance is either 0.78, 0.70, 0.94, or 0.86 mm with
the first being the most predominant. Of the 25 scans in the
training set, 13 scans have slice distance 0.78 mm and of the
114 scans in the test set, 72 have that same slice distance. For
these 72 scans, the DSC of the medial cartilage compartments
SD. For the other resolutions in the test set the
average DSC is
SD. Of these remaining scans, 32
0.70 mm ,
B. Correlation Between the Volume and Area Estimate and
Typical quantitative disease markers for OA is the articular
have been dedicated to evaluation of them –. In this
study, we evaluate the volume and surface area estimates ob-
tained directly from the automatic segmentation. The volume
estimate is directly obtainable by summing all voxels classified
as cartilage, and an estimate for the surface area is obtained by
creating an isosurface using a smoothed version of the binary
segmentation. But a voxel based method alone does not allow
for morphometric quantification, and for measuring the thick-
ness, we fit a deformable shape model to the cartilage so that
FOLKESSON et al.: SEGMENTING ARTICULAR CARTILAGE AUTOMATICALLY113
P-VALUES FOR T-TESTS OF SEPARATING GROUPS USING THE VOLUME
ESTIMATES. P1 IS THE P-VALUE FOR SEPARATION OF HEALTHY ???? ? ??
FROM BORDERLINE TO OA ???? ? ??, AND P2 IS SEPARATION OF HEALTHY
AND BORDERLINE ???? ? ?? FROM CLEAR OA CASES ???? ? ??.
M STANDS FOR MANUAL SEGMENTATIONS AND PA ARE VALUES FROM
AUTOMATIC SEGMENTATION AFTER POSITION ADJUSTMENT
P-VALUES FOR T-TESTS OF SEPARATING GROUPS USING THE AREA ESTIMATES.
P1 IS THE P-VALUE FOR SEPARATION OF HEALTHY ??? ? ?? FROM
BORDERLINE TO OA ??? ? ??, AND P2 IS SEPARATION OF HEALTHY AND
BORDERLINE ??? ? ?? FROM CLEAR OA CASES ??? ? ??. M STANDS
FOR MANUAL SEGMENTATIONS AND PA ARE VALUES FROM AUTOMATIC
SEGMENTATION AFTER POSITION ADJUSTMENT
thickness can be measured through the normal direction of the
cartilage surface at anatomical well-defined locations. This is
however not within the scope of this paper, for thickness mea-
surements of the data set, see .
populations of the volume and area estimates using an unpaired
students -test. The results are displayed in Tables III and IV,
and since knees with
uate populations both by including these cases to the healthy
population and to the OA population. It can be seen that for the
volume estimate the most confident separations occurs for tibial
cartilage, and for the area estimate statistical significant separa-
tion is obtainable only from tibial cartilage.
Since our test subjects come in all shapes and sizes, we
normalize the volume by the width of the tibial plateau cubed
and the surface area by the tibial plateau width squared. In
Figs. 5 and 6, the normalized volume and surface area estimates
for medial tibial and femoral cartilage together are plotted
are borderline cases we eval-
In this paper, we have presented a fully automatic framework
for segmentation and quantitative assessment of the articular
cartilage in the knee. This is, to our knowledge, the only fully
and agreement with manual segmentations and is evaluated on
a fairly large data set (139 scans) consisting of both healthy and
osteoarthritic test subjects.
Fig. 5. Separation between different OA populations using the KLi and the
normalized tibial medial cartilage volume from automatic (position adjusted)
malized tibial medial cartilage surface area from automatic (position adjusted)
Robustness against the inevitable problem of changes in test
subject placement in the scanners is obtained with an iterative
scheme, which facilitates low interscan variability of the carti-
The medial tibial cartilage gives the best inter-scan repro-
ducibility with mean absolute difference of 5.8% and 4.3% for
the volume and area estimates, and separation between popu-
lation with -values of 0.003 and 0.005 for separation between
Fat suppression and high-field magnets significantly improve
image quality with better contrast between tissues and higher
resolution. Since our method compares well to manual segmen-
tations using the lower quality images from a low-field scanner,
we can hope that the method will perform at least as well on
high-field fat suppressed MRI, assuming we would have ac-
cess to a similar amount of training data. Future work will in-
volve evaluating the method on high-field data. Our segmenta-
tion method can handle images with somewhat different reso-
lution, however, it is possible and remains to be investigated if
features present at higher resolutions can advance the results.
114IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 26, NO. 1, JANUARY 2007
By using binary classifiers we not only avoid the problem of
also established a framework for multi-class classification that
in the future can be extended to incorporate all cartilage com-
partments by incorporating binary classifiers trained separately
for the remaining compartments.
Our method is trained and evaluated on low-field MRI, and
even though there is no well established accuracy validation for
low-field MRI, we show that statistically significant differences
between healthy and osteoarthritic populations are detectable
our method combined with low-field MRI data may be useful
in clinical studies, particularly multicenter clinical studies since
the method is completely automatic, has high reproducibility,
and is robust to changes in knee placement in scanner.
 D. Jackson, T. Simon, and H. Aberman, “Symptomatic articular
cartilage degeneration: The impact in the new millenium,” Clin.
Orthopaedics Related Res., vol. 391, pp. 14–25, 2001.
 D. T. Felson, R. C. Lawrence, M. C. Hochberg, T. McAlindon, P. A.
Dieppe, M. A. Minor, S. N. Blair, B. M. Berman, J. F. Fries, M. Wein-
berger,K. R. Lorig, J. J. Jacobs, and V. Goldberg, “Osteoarthritis: New
insights, part 2: Treatment approaches,” Ann. Int. Med., vol. 133, no. 7,
pp. 726–737, Nov. 2000.
 H. Graichen, R. Eisenhart-Rothe, T. Vogl, K.-H. Englmeier, and F.
Eckstein, “Quantitative assessment of cartilage status in osteoarthritis
by quantitative magnetic resonance imaging,” Arthritis Rheumatism,
vol. 50, no. 3, pp. 811–816, Mar. 2004.
 E. Pessis, J.-L. Drape, P. Ravaud, A. Chevrot, and M. D. X. Ayral,
“Assessment of progression in knee osteoarthritis: Results of a 1 year
study comparing arthroscopy and mri,” Osteoarthritis Cartilage, vol.
11, pp. 361–369, 2003.
 C. Peterfy, G. Gold, F. Eckstein, F. Cicuttini, B. Dardzinski, and R.
Stevens, “Mri protocols for whole-organ assessment of the knee in
osteoarthritis,” Osteoartritis Cartilage, vol. 14, supplement A, pp.
 B. Ejbjerg, E. Narvestad, S. J. adn H. S. Thomsen, and M. Ostergaard,
“Optimised, low cost, low field dedicated extremity mri is highly
specific and sensitive for synovitis and bone erosions in rheumatoid
arthritis wrist and finger joints: A comparison with conventional
high-field mri and radiography,” Ann. Rheumatic Diseases, vol. 13,
articular cartilage lesions: Experimental evaluation of low- and high-
field-strength mr imaging at 0.18 and 1.0 t,” J. Magnetic Resonance
Imag., vol. 11, pp. 678–685, 2000.
 B. Kladny, K. Gluckert, B. Swoboda, W. Beyer, and G. Weseloh,
“Comparison of low-field (0.2 tesla) and high-field (1.5 tesla) mag-
netic resonance imaging of the knee joint,” Arch. Orthopaedic Trauma
Surg., vol. 114, no. 5, pp. 281–286, 1995.
 B. Kersting-Sommerhoff, P. Gerhardt, W. Golder, N. Hof, K. Riel, H.
Helmberger, M. Lentz, and K. Lehner, “Mri of the knee joint: First
results of a comparison of 0.2-t specialized system and 1.5-t high field
strength magnet,” Fortschr. Rontgenstr., vol. 162, no. 5, pp. 390–395,
“0.2-tesla magnetic resonance imaging of internal lesions of the knee
joint: A prospective arthroscopically controlled clinical study,” Knee
Surg. Sports Traumatol. Arthrosc., vol. 7, pp. 37–41, 1999.
 R. W. Huegli, P. F. Tirman, H. M. Bonel, H. Staedele, S. Zaim, M.
Grigorian, and H. K. Genant, “Use of the modified three-point dixon
ages on an open magnet,” Eur. J. Radiol., vol. 11, no. 7, pp. 473–474,
 T. Stammberger, F. Eckstein, M. Michaelis, K.-H. Englmeier, and M.
Reiser, “Interobserver reproducibility of quantitative cartilage mea-
surements: Comparison of b-spline snakes and manual segmentation,”
Magn. Resonance Imag., vol. 17, no. 7, pp. 1033–1042, 1999.
 J. A. Lynch, S. Zaim, J. Zhao, A. Stork, C. G. Peterfy, and H. K.
Genant, “Cartilage segmentation of 3-D mri scans of the osteoarthritic
Imag. 2000: Image Process., vol. 3979, pp. 925–935, 2000.
“Automatic measurement of subtle changes in articular cartilage from
mri of the knee by combining 3-D image registration and segmenta-
tion,” Proc. SPIE Med. Imag. 2001: Image Process., vol. 4322, pp.
 S. Solloway, C. Hutchinson, J. Vaterton, and C. Taylor, “The use of
active shape models for making thickness measurements of articular
cartilage from mr images,” Magnetic Resonance Med., vol. 37, pp.
 V. Grau, A. Mewes, M. Alcaiz, R. Kikinis, and S. Warfield, “Improved
watershed transform for medical image segmentation using prior in-
formation,” IEEE Trans. Med. Imag., vol. 23, no. 4, pp. 447–458, Apr.
 S. K. Pakin, J. G. Tamez-Pena, S. Totterman, and K. J. Parker, “Seg-
mentation, surface extraction and thickness computation of articular
cartilage,” Proc. SPIE Med. Imag. 2002: Image Process., vol. 4684,
pp. 155–166, 2002.
 J. G. Tamez-Pena, M. Barbu-McInnis, and S. Totterman, “Knee car-
tilage extraction and bone-cartilage interface analysis from 3-D mri
data sets,” Proc. SPIE Med. Imag. 2004: Image Process., vol. 5370,
pp. 1774–1784, 2004.
 S. K. Warfield, M. Kaus, F. A. Jolesz, and R. Kikinis, “Adaptive,
template moderated, spatially varying statistical classification,” Med.
Image Anal., no. 4, pp. 43–55, 2000.
 S. K. Warfield, C. Winalski, F. A. Jolesz, and R. Kikinis, “Automatic
segmentation of mri of the knee,” presented at the ISMRM 6th Sci.
Meeting Exhibition, Sydney, Australia, Jul. 1998.
 K. Li, S. Millington, X. Wu, D. Z. Chen, and M. Sonka, “Simulta-
neous segmentation of multiple closed surfaces using optimal graph
searching,” in Information Processing in Medical Imaging: 19th Inter-
national Conference.New York: Springer, 2005, vol. 3565, Lecture
Notes in Computer Science.
 S. Millington, K. Li, X. Wu, S. Hurwitz, and M. Sonka, “Automated si-
multaneous 3-D segmentation of multiple cartilage surfaces using op-
timal graph searching on mri images,” Osteoarthritis Cartilage, vol.
 T. Dunn, Y. Lu, H. Jin, M. Ries, and S. Majumdar, “T2 relaxation
time of cartilage at mr imaging: Comparison with severity of knee os-
teoarthritis,” Radiology, vol. 232, no. 2, pp. 592–598, 2004.
minedby mri,”OsteoarthritisCartilage,vol. 14, pp. 1781–1786,2004.
 J. Kellgren and J. Lawrence, “Radiological assessment of osteo-
arthrosis,” Ann. Rheumatic Diseases, vol. 16, no. 4, 1957.
 S. Arya, D. Mount, N. Netanyahu, R. Silverman, and A. Wu, “An op-
timalalgorithmfor approximatenearest neighbor searchingin fixeddi-
 J. Folkesson, O. F. Olsen, P. Pettersen, E. Dam, and C. Christiansen,
“Combining binary classifiers for automatic cartilage segmentation in
knee mri,” in ICCV 1st Int. Workshop: Comput. Vision Biomed. Imag.
Appl., 2005, pp. 230–239.
 J. Folkesson, E. Dam, O. F. Olsen, P. Pettersen, and C. Christiansen,
“Automatic segmentation of the articular cartilage in knee mri using a
hierarchical multi-class classification scheme,” in 8th Int. Conf. Med.
Image Comput. Comput.-Assist. Intervention (MICCAI’05), Palm
Springs, CA, 2005, pp. 327–334.
 C.-H. Yeang, S. Ramaswamy, P. Tamayo, S. Mukherjee, R. M. Rifkin,
M. Angelo, M. Reich, E. Lander, J. Mesirov, and T. Golub, “Molecular
 K. H. Zou, S. K. Warfield, A. Bharatha, C. M. Tempany, M. R. Kaus,
S. J. Haker, W. M. W. III, F. A. Jolesz, and R. Kikinis, “Statistical
validation of image segmentation quality based on a spatial overlap
index,” Acad. Radiol., vol. 11, pp. 178–189, 2004.
IEEE Trans. Pattern Anal. Mach. Intell., vol. 22, no. 1, pp. 4–37, Jan.
 J. J. Koenderink, “The structure of images,” Biol. Cybern., vol. 50, pp.
 L. Florack, “The syntactical structure of scalar images,” Ph.D. disser-
tation, Univ. Utrecht, Utrecht, The Netherlands, Oct. 1993.
 J. Blom, “Topological and geometrical aspects of image structure,”
Ph.D. dissertation, Utrecht Univ., Utrecht, The Netherlands, 1992.
FOLKESSON et al.: SEGMENTING ARTICULAR CARTILAGE AUTOMATICALLY 115 Download full-text
 M. Descoteaux, L. Collins, and K. Siddiqi, “Geometric flows for seg-
menting vasulature in mri: Theory and validation,” in 7th Int. Conf.
Med. Image Comput. Comput.-Assisted Int. (MICCAI’04), St Malo,
France, 2004, vol. 3216, pp. 500–507.
 J. Weickert, Anisotropic Diffusion in Image Processing.
Germany: Teubner-Verlag, 1998.
 E. B.Dam,J.Folkesson,M.Loog,P.C.Pettersen,andC.Christiansen,
“Efficient automatic cartilage segmentation,” in MICCAI Joint Disease
Workshop, Copenhagen, Denmark, 2006 [Online]. Available: http://
 J. Folkesson, E. B. Dam, O. F. Olsen, P. C. Pettersen, and C. Chris-
tiansen, “Position normalization in automatic cartilage segmentation,”
in MICCAI Joint Disease Workshop, Copenhagen, Denmark, 2006
[Online]. Available: http://www.itu.dk/image/joint
 R. Burgkart, C. Glaser, A. Hyhlik-Durr, K.-H. Englmeier, M. Reiser,
and F. Eckstein, “Magnetic resonance imaging-based assessment of
cartilage loss in severe osteoarthritis,” Arthritis Rheumatism, vol. 44,
no. 9, pp. 2072–2077, Sept. 2001.
 T. Stammberger, F. Eckstein, K.-H. Englmeier, and M. Reiser, “Deter-
mination of 3-D cartilage thickness data from mr imaging: Compputa-
tional method and reproducibility in the living,” Magnetic Resonance
Med., vol. 41, pp. 529–536, 1999.
 J. Hohe, G. Ateshian, M. Reiser, K.-H. Englmeier, and F. Eckstein,
“Surface size, curvature analysis, and assessment of knee joint incon-
gruity with mri in vivo,” Magnetic Resonance in Medicine, no. 47, pp.
 E. B. Dam, J. Folkesson, P. C. Pettersen, and C. Christiansen, “Au-
tomatic cartilage thickness quantification using a statistical shape
model,” in MICCAI Joint Disease Workshop, Copenhagen, Denmark,
2006 [Online]. Available: http://www.itu.dk/image/joint